Hi HN. I wrote this post after getting frustrated by the lack of ways to run the new Gemma 4 Drafter models, and mainstream tools not prioritizing this, and hiding all the performance levers.
I ended up getting a modern 26B MoE model (Gemma 4) running at reading speed on an old recycled server with a single Xeon E5-2620 v4 and 128GB of DDR3 RAM (and no GPU). It took a lot of work, but it actually worked out somehow.
I've also linked the quants at the end, but they're not gonna run unless you use the ik_llama-cpp fork I mention, see other posts for more details.
I'm not an ML engineer, so I'm by no means an expert, and the server is busy acting as a Nix cache, but if you have any question, I can try to answer, but best effort.
"-t 8 matches physical cores. The machine has 16 SMT threads but only 8 cores. On a memory-bound workload, oversubscribing threads adds scheduling cost without adding throughput: the cores are waiting on DDR3, not on each other."
But ... isnt that a classic use case for SMT? Giving T1 sth. to do while T0 is waiting on DDR(3) and vise-versa?
I also dont understand the explanation of "--cpu-moe".
If an expert has ~ 4.0 GiB of Parameters, why does optimizing the sequence of experts minimize cash trashing? With 20 MiB of L3 Cash vs 4.0 GiB of Parameters, it wont cash any noticeable amount of the Parameters, will it?
As mentioned by others, only some Intel Xeon E5-2xxx v4 did support DDR3, and according to Intel, the E5-2620 v4 is not one of them.
> But ... isnt that a classic use case for SMT? Giving T1 sth. to do while T0 is waiting on DDR(3) and vise-versa?
Waiting in terms of latency. When the bus is mostly empty and it takes a while to make a round trip it's great to try to find a few extra passengers to put on it. When the buses are all completely full adding the extra riders just makes the bus stop that much more chaotic.
This is ironically a pretty solid use case for (ex VLIW research) ILP-optimizing compilers.
Given knowable runtime hardware usage patterns (huge bursts of memory bandwidth saturation) and a single limited core/thread-shared resource (memory bandwidth), one could optimize for the constraint ahead of runtime.
Because most of the performance optimization levers you have available to pull are (a) trade compute for memory bandwidth (e.g. compression), (b) preload when memory bandwidth is available, (c) optimize the choice of what's in cache when, (d) align to cache size / memory boundaries.
Or tl;dr, try to approximate GPU ISAs at the CPU compiler level. (Which why would anyone but hobbyists, because everyone else just buys pallets of Nvidia/AMD or designs their own ML chips?)
Something doesn't add up here. As someone who has only recently built a home-server from an E5-26xx v2 on DDR3 RAM (because I have a sh*tload of 32g DDR3 DIMMs), I can confidently say that the newer cores (E5-26xx v3 and v4) only run on DDR4 memory...
So either you have a v2 instead of a v4 (and run on DDR3 memory), or you have a v4 but with DDR4 memory (not DDR3)
There are some OEM-only v3/v4 parts with dual memory controllers (because of a RAM supply crunch at the time, funnily enough), but the E5-2620 v4 is not one of them. The classic example is the very popular 12-core E5-2678 v3.
It looks like Supermicro had some DDR3 Xeon v3/v4 boards, and the first thing that came to mind was a Shenzen workstation/gaming board using recycled parts... haven't searched on that but it's bound to exist.
> So either you have a v2 instead of a v4 (and run on DDR3 memory), or you have a v4 but with DDR4 memory (not DDR3)
Yup that's odd... I've got a Xeon 2680 v4 (14 cores) (amazing bargain of a little beast btw) and it's indeed on DDR4 and I saw all Xeons v4 as supporting DDR4 only.
Full spec (brand/model/mobo type) would have been nice: mine's an HP Z440 workstation repurposed as a server (which I only turn on when I'm working and which I religiously turn off before going to bed).
Noted, and agree (it looks like it has also already been clicked, which I dislike). I honestly I need to redo the themes.
> You say it runs "at reading speed". Have you benchmarked it?
At some point a few weeks ago, yes I think so, but I didn't write it down for some reason... so I'll have to find a time when it's not busy and do it again without a noisy system. Right now the system is noisy, but that said doing it like this:
llama_print_timings: load time = 83911.65 ms
llama_print_timings: sample time = 26.99 ms / 128 runs ( 0.21 ms per token, 4742.15 tokens per second)
llama_print_timings: prompt eval time = 343.41 ms / 7 tokens ( 49.06 ms per token, 20.38 tokens per second)
llama_print_timings: eval time = 10639.36 ms / 127 runs ( 83.77 ms per token, 11.94 tokens per second)
llama_print_timings: total time = 11114.98 ms / 134 tokens
So 11.94 tokens per second while it's also playing binary cache and CI builder.
When I do it properly, I'll add it to the blog as well!
And if you ever run out of things to do in your copious free time, it looks like that PR #1744 was merged without the has_target_ctx assert two days after you uploaded your drafter quants. So you can now redo all your quants and rerun all your benchmarks ;-).
llama-bench is part of the llama-cpp package, but from recent experimentation, the settings it is able to (or is documented to?) accept lag behind somewhat. Not sure whether it would accept all of the esoteric settings in the article?
I'm pretty sure eval time is token generation time where it's actually outputting new tokens. If you're getting a thousand per second on that, I'd love to know on what.
From the prompt timings above, it seems like 'prompt eval time' is the equivalent to 'processing time for input tokens'.
Hyperscalers can perform this evaluation very quickly because evaluation can be significantly parallelized. The layer `i` output of token `j` only requires access to the layer `i-1` output of all previous tokens, so a parallel frontier develops. Token (0,0) [(token, layer)] is processed first, then tokens (0,1) and (1,0) can be processed in parallel, then (0,2), (1,1), and (2,0), and so on.
The maximum parallel width becomes equal to the number of layers in the model. Gemma 4 26B-A4B model discussed in this article evidently has 30 layers, giving a 30-fold speedup if the system were otherwise unconstrained (all layers can be run in parallel, and one full set of layer outputs is completed in the KV pass for each pass of the parallel sweep).
In the specific output above, however, the input prompt is only seven tokens long so there are probably considerable non-amortized spinup effects at play.
Seven tokens long input isn't very realistic, is it? For coding tasks it's normal for the input to be thousands or 10s of thousands. If it wasn't for prefix caching it'd be one miserable experience, but even then at the very best the input is often in hundreds each time. And don't even try to dump some logs into the prompt.
> Seven tokens long input isn't very realistic, is it?
The test prompt above was "Why is the sky blue?", so there's the seven tokens. I meant to highlight that because I'd expect processing of a thousand-token input to be faster per token than presented.
How many watts is that setup? Cool you got it to work, but maybe only useful for vintage / retro computing rather than practical if the energy consumption makes it economically wasteful.
How many kWh to fabricate a brand new machine better suited to the task?
As long as performance is useable (apply your own metrics!), pulling it from existing hardware is likely the option with the lower eco footprint.
Also: chances are it'll only be used for this purpose occasionally, and/or for a short while. In that scenario [fabricating new hardware] always has the bigger eco footprint.
I don’t know why you’d assume that an older system is lower footprint.
If you’ve got something consuming 100 watts average over your 24 hour period, and your electricity costs 20 cents per kWh, you’re already spending almost as much as a Claude subscription.
Just on electricity, this assumes your hardware never fails and you never incur any additional costs.
There’s a big reason why newer more efficient hardware is in demand. Something that’s 10+ years old has drastically worse performance per watt.
Obviously I am not saying to throw away your old hardware as a rule but there is a point where some of this old stuff just isn’t even worth running.
I have two LARGE Xeon systems of this era that I used to use when I was heavily involved with Kubernetes and needed to build out a home lab. One is 2x Xeon w/ 256 GB of ram, and one is 1x Xeon w/ 512GB of ram. Both are slow as dogs, and both of them take up at least 150+ watts with only one power supply. My 12th gen Intel Nuc is so, so much faster and efficient. I'm recycling the Xeon systems.
Would you consider improving the website's layout? Right now I find it below average quality and very distracting. Whether you are an engineer or not is not really important; great engineers can write horrible text or use a layout that is not ideal, for instance.
We’re not there yet, but the obvious endgame of the present bubble insanity is open models running on local hardware and devices are “good enough” for most use cases. That will completely implode what’s going on at the moment in tech.
Happened to me. CoPilot changing prices prompted me to cancel my CoPilot subscription and install a local coding model running entirely in VRAM. Will call Claude APIs when I get really stuck, but I should be able to handle 80% of my needs with a dumber local model.
For a long time, too. Programming languages rarely change much, techniques rarely change, so I should be able to use said model for I hope at least five years; and if at any time they optimize local models to cram even more intelligence into the same amount of VRAM, I can upgrade to that.
You just described the absolute nightmare scenario for the newly minted trillion-dollar companies whose only hope is for enterprises and SMB to move all their business processes to the cloud, with employees competing at token maxxing.
This. OpenAI and Anthropic are ultimately compute infrastructure plays and not really AI. Everyone will have models, they'll have the ability to run them. This is why the GPU shortage is in their favor.
And like Google and Meta, these companies are going to morph into advertising giants. Advertising is an economic black hole and it eats everything that comes close.
How does that view align with Anthropic leasing data centers from others?
I don’t know OpenAI’s infra, but to the extent they are buying GPUs and building data centers with their own money, that sounds like a bad move.
Satya has mismanaged the AI transition in many ways, but one thing he got right is that models are commodities, and the value is in applications that apply them to create user benefit. I agree that any company trying to build a moat with a model is not long for this world.
Do you think there will still be an incentive to release weights in that scenario? Everyone will have models only if there continue to be companies releasing weights.
Companies won't but I suspect this is a role that something else open source-y will fill that niche. Maybe orgs like wikimedia or internet archive, maybe some hackers just making things, maybe nation states that want to disrupt other players. Also model training will get better and better both on the algo and the hardware side. You can easily see a world where you might be able to train a good enough model on a home lab in a few days.
I wouldn't say "completely implode", too much money was poured int it, but it's clear we're heading in that direction. You get a model that is "good enough", plus privacy, plus savings in the long term.
Paradoxically, the better results we get from general harness of coding agents, the less moat Claude and co. get. It's unbelievably how fast some open models outpaced frontier models of just a few months ago.
I disagree. We are currently in a weird period where these frontier AI companies are losing tons of money even on the subscription-based AI models. It's just too compute intensive and there's no way most people are going to be buying the kind of hardware required to run $20 worth of inference every day.
Sadly - it's going to be ads. Advertising is going to get in there and enshittify the whole thing because as always, advertising income is too easy and too plentiful for any company to resist.
Right now the models are fairly agnostic, but we are a hair-breadth away from ChatGPT responding with, "the right tool for this job is a circular saw - something like the Milwaulkee M18, which happens to be on sale at Home Depot this weekend."
$20/day x 250 days per year x # devs/agents/etc = $$$. About $5k per dev at that daily use case.
Enough to validate repurposing an existing workstation with enough RAM, or finding a used high VRAM GPU, or in my case buying a Strix Halo system for home lab and local models.
The future is once again not cloud based, for AI tools.
Most people are running a whole lot less than $20's worth of tokens per day on cloud platforms. (Is that assuming a frontier model? 1M output tokens per day?) Local hardware could easily take up that workload, at least the part of it that's non-time-critical.
The advertising future looks like that to me, too. Service proxies like OpenRouter might talk about price optimization, maybe some ad filtering. But I expect proxies will have malicious entries, too, surreptitiously altering agentic prompts.
Ads are usually the workaround where you don’t deliver enough value to get people to subscribe or payments are unavailable for some reason.
It makes sense to show some ads and get some money at low volume (like a faraway reader wanting to read a story in your local newspaper) but taking money from regular users directly will pay much more.
Newspapers are happy to cannibalize 99% of their ad revenue with a paywall if that 1% subscribes because that’s how much more money you make from someone paying $10-$20/month vs ads.
But yeah, if people use it as a buying recommendation engine, that’s where the money is on ads/referrals but a lot of AI use has little/no connection to buying intent touchpoints.
Newspapers had no choice after craigslist and later Google/Facebook took all their classified revenue.
LLMs may or may not be able to cover their costs with it. We'll see - I suspect product placement as recommendations will become a thing as it won't take as much GPU to give a "recommendation" on "the best widget for X". I firmly expect it to become enshittified the same way google and amazon search has.
I find that hard to believe. The AI companies will want to control what's possible and find new things to do that "need" their services. Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
> Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
I think you've misunderstood what good enough means in the context - which is a model capable of completing the tasks assigned to it without having the breadth of full generalization. Your analogy breaks down because of this - we did get 'good enough' spec profiles for different hardware. That thing you're wearing on your wrist won't have the same specifications as the box you use to play games.
I think you've misunderstood the analogy. Just ignore it, analogies mostly break down anyways.
> a model capable of completing the tasks assigned to it
The thing is, the "task assigned to it" is changing with improved capabilities. If everyone around you in 2036 is using general AI to do amazing stuff, you will probably have little interest in vibe coding slop like it's 2026.
>Otherwise it would be like Intel and Microsoft had decided in the year 2000 that computers are "good enough" now and we would have explored what's possible with that hardware ever since.
That would be the dream... no fucking Electron! No lockdown modules.
> The AI companies will want to control what's possible and find new things to do that "need" their services.
That's correct. The problem is they have smart people, tons of money, and several years to figure that out, and the best thing they can come up is a coding agent.
I run my word processing software on my apple 2 (a total joke of a computer) instead of running it on the WANG.
I run my book keeping software on visicalc instead of the IBM.
I run my simulation software on my IBM PC (I even paid for the 8087!) instead of the VAX.
Moore's law has, at least so far, allowed the pioneers with toy computers to grow their toys big enough to solve "big boy" problems after some time has allowed the toy computers to be faster and the pioneers have scaled their crappy home-grown solution to solve their 60% of the problem that was originally solved by some enormous complex system.
Eventually the toy infrastructure gets expensive and solves 90-120% of the "big iron" problem space, but it also grows to cost as much as the big iron solution, but then a new generation of toy software and toy systems emerges to disrupt the "big iron" systems.
Under appreciated requirement for this to work in post-cloud times: open source
If a vendor can SaaS a solution, then enterprise is generally happy (they don't want to have to hire folks for maintenance), and that completely locks out any ability to run locally.
Between enterprise's ambivalence and the obvious financial incentive to vendors, you get SaaS-only products.
It's a huge difference. If you had AI sufficiently good running locally on a phone, you could devise workflows for things like basic digital hygiene, technical assistance, and tedious tasks like inbox management, image sorting, device updates, and so on. Privacy and security gets a big boost past some local competence threshold, and we're nearly there.
Make the local AI competent enough to do good image generation and editing, realtime voice and music generation, handle agentic tasks with a framework like Hermes, and you can take your AI places to do tasks in contexts that are inaccessible to or inappropriate for cloud.
Frontier big platform models will be the best, but there's a level of "good enough" for local uses that we're already seeing flourish, and "good enough" for the average joe is almost here.
Phones and laptops are terrible devices for local AI, way too constrained by bad thermals and small batteries. MiniPC's (many of them using mobile hardware) don't have that particular issue, and can easily run on a 24/7 basis.
It's a little different because cloud and blogs didn't actively get in the way of your home compute. To wit, the various cost spikes for hardware.
People -- WANT -- this technology on their home devices and (apparently?) the providers of this tech don't seem to be running a profit so they probably don't want the maintenance tail on their side either.
I think it's a bit different. Inevitable that this becomes a household-run thing? Not likely.
The primary feature of a blog or any website is that it is available around the clock, that is the primary feature of cloud: around on the clock computer and network that scales on demand.
The primary feature of "AI" is to process information and reason with a natural language interface at speed, the primary feature of AI bigboys is to provide the machinery that runs the "models".
Running an LLM locally is theoretically viable. Running your blog on your laptop is never viable (unless you hook it up like a server). One just requires compute while the other a stable network.
Gamers Nexus has a good video on this, but if NVIDIA exits the consumer market, and honestly why would they stay when they can charge up to a 100x for the same wafer space for enterprise, AMD would likely do the same.
Only Apple really makes consumer hardware suitable for running things locally then, and maybe some weird Qualcomm ARM chip for Windows.
It will be hard running things locally if nobody is supplying the hardware.
Nice post and technically impressive work. I agree we need to understand the build pipeline and be able to do things locally. However, depending on your electricity cost, it might not make sense financially. These old servers are not energy efficient at all (I'm guessing that old Xeon server will easily pull 200W on load), and that model is currently at 0.1$/0.3$ per 1M tokens (with 76 tps and 262k context) in Openrouter (also, these servers are LOUD).
EDIT: I stand corrected, 200W is apparently way too high of an estimate. I used to run a bunch of old Xeon servers and they slurped watts like crazy, but I can't remember which ones exactly those were.
OK, then you're in luck. I had a bunch of old 1U rack servers and even in the next room it was too annoying to run them (they had a bunch of 40mm fans which always ran at full speed, because in a server room, no one can hear you scream).
Could it just be really bad cooling? Looking at 9800X3D, it seems like it's running in a similar range wrt TDP unless you really push the 9800X3D. I'm comparing with desktop cpu's because that's what my workload is. cpu governor is set to performance (no schedutil). No audible change in fan speed during heavy compilation or gaming (very silent humming), and i don't have any fans beside cheap intake, cpu and exhaust fans (1 each) + an excessive amount of dust.
These servers had no fan control whatsoever, they always ran full blast. That's not untypical for rack servers, because as written: they are designed for server rooms, and you're supposed to wear ear protection there anyway... Yes, I could've modified them, but I ditched them because running them simply made no sense (especially the high idle power consumption was ridiculous).
Only when you remove it from the original server or enable low fan mode (if available). Most 1U/2U cases will happily blow at full speed well over 90db.
You likely need to replace the flow-through server chassis system with an active "normal" cooler to achieve a bit of silence.
85W might be about right. My old server CPU is in the same ballpark and compiling kernels it reached about 90w in power usage. If you want to keep it running: idle is not very low power unless you have one of the "low power" L versions, keep that in mind.
Get a 4U case, many options if you want to combine it with a NAS. Not hard to cool and keep somewhat quiet. If you can store it in a closet or something that helps too.
Well, you can use it for lots of other things as well.
Compared to the cloud you can probably save up to buy a new server every month. And don't underestimate the gains of having something to experiment on and play with.
These servers are loud if you're trying to fit them into a 1U or 2U, which requires high speed fans to generate the necessary static pressure to push air through the case. I run a similar setup in a 4U case with slow 120mm fans and it's fine.
Glad to see other people realizing this. I've been running Gemma 26B-A4B Q4 on a 2012 Xeon with 16GB to 24GB of RAM in a container. It's getting around 8 to 12 tokens per second. Obviously it's not comparable to huge contexts and running it on a GPU and the image decoder in llama.cpp is super slow compared to a GPU but for some small automation tasks and general trivia questions it's decent. The speed is just enough to not have to wait for it to finish so you can read along.
Here's my setup. You may want to figure out what the best optimizations are for your specific CPU like AVX2 because mine didn't have most of them. I did try MTP briefly but I wasn't getting performance improvements. You could play around with the batch sizes for cache or context or go even lower for Q2 and don't overcommit on threads either, but I would suggest either defaults or trying out llama-bench. This isn't by any means the best I assume but it worked decently for me and I sometimes swap out Gemma for Qwen. You could also lower q8_0 to q4_0 for more context but it could hurt quality some say, altough I have noticed it too on some models.
Result is ~12 tokens per second, as reported by OP down in these comments here.
An impressive effort, and better than I would have thought possible on this hardware -- but still pretty far short of what one needs for an satisfactory interactive session.
Especially if you consider those smaller models are really cheap and fast on platforms like openrouter. Often by the factor 100-500 cheaper than SOTA models, and 2-5x in TPS.
What intrigues me the most about AI progress, is not AGI or the model du jour by $AI_UNICORN, but rather what can be run locally. I remember having an amusing, but rather useless model in a beefy gaming PC that I had 6 years ago; and now, something that’s a hundred times better on my M5 laptop.
Should the market react to the memory shortage, the progress of the Apple silicon continue at the same pace, and what we’ll be able to run locally in 6 years will be very exciting. or frightening.
Also I don’t know what this means for the valuation of the AI companies. I remember asking about this very idea to one of their employees at an event and instead of answering he bailed out to grab a cocktail.
- There is no "moat" (lasting, easy-to-defend technological edge) in AI model businesses. There are just short-term advantages.
- An AI business is a capital-intensive business, just like old factories. Data centers are expensive, models are energy-hungry, and the hardware inside must be replaced every 3–4 years.
- Smaller, specialized models eat margins from below. Transcription, voice, or image detection do not need large models.
There is no reason to expect high margins like you can in traditional software business. Benefits of AI go mostly to consumers.
edit: There is potential for economies of scale. Few megacorps can strive for cost advantage when they achieve scale (Microsoft, Google, Amazon and Meta)
It does seem like the structural characteristics we’ve observed so far suggest there is a kind of flywheel from short-term to long-term advantage due to the capital requirements at various levels.
If you’re Nvidia, making the best GPUs today, the expanding wavefront of demand is consuming them with volume and margins to give you a huge edge in building out the best next generation of GPUs. Similar to how the mobile wave gave TSMC sustained advantage for about a decade now.
I’m guessing this is also what we’re seeing as Anthropic and OpenAI swap spots in the token-vendor market.
What you can run locally in consumer hardware is progressing pretty well.
If you get a not-quite-the-best gaming GPU like a 5080, you can run local models that are better than the state of the art from early 2025. Depending on what you want to do, you might have to switch models. The one size fits all huge models are still a data center thing.
--what this means for the valuation of the AI companies
Probably nothing. Most users have no idea what an LLM is or how it runs. Anecdotally speaking, I see many LLM users default to whatever their day job provides to them. And even slightly more sophisticated users seem ok with paying for their openai or anthropic subscriptions.
Maybe we will see a small but dedicated group of open weight model users who prefer local llm, but everybody else will just consume from the big providers? The scenario might look something like OS choices today - a small, committed group of Linux users vs the vast majority of other users running Windows, MacOS, or Chrome?
Its a convenience thing. You can run a whole lot of stuff locally from wikipedia to social media/email/video servers whatever. Most people with a full time job and 2 kids dont do it cause who has time and energy to patch and maintain the ever growing complexity of this stuff. These systems will keep growing complex. That also means more bugs. Age old tradeoff between freedom and convenience.
You can run mediawiki at home but you won't have wikipedia. You can run a video server but you won't have all the movies that Netfix has. A local model is actually the real thing.
This has always been true of software, particularly games. You can get a 5-6 year old game for a fraction of the price, and run it on modest hardware. But the industry wont sit on its hands for 5 years, there will be newer software that requires better hardware.
A new game is a totally new world with everything created from scratch. A creation. A model, on the other hand, is a reinterpretation machine for hundreds of years of human creations, but not a creation in itself, more like a discovery.
You would think that by now we would have a much better Bitcoin that's taking over the payment networks of the world but what we actually got is a shitload of shitcoin.
The E5-2620 v4 is great. Have been using it for 10 years now. Wanted to upgrade until I saw current prices. I have 64 GB ddr4. Paired it with rx 9060 xt 16 GB and games run as fast as ever. Perhaps the cpu is a slight bottleneck in DOOM The Dark Ages, but i'm at 60 fps, so no problem. Light llm on the gpu is a nobrainer, and it's cool to see that things can be tuned to run ok on the cpu. I bought 2667 v4 a month ago for 30$. I'd expect it to give a decent performance boost but I just haven't had the need for it yet, but pushing into llm like in the article I'd probably upgrade because 2667 can handle slightly faster ram.
> The E5-2620 v4 is great. Have been using it for 10 years now.
10 years? Damn, that is a long time. I always assumed that heat-induced damage will kill a CPU after a certain amount of time (5-7 years). Am I wrong here? I assume yes. Or are CPUs must stronger/tougher than the bad old days?
This is among the "real" differences between workstation/server CPUs and commodity chips for laptops/desktops/handhelds.
Even then, if a commodity chip isn't pushed full tilt at all times, and assuming that the venting and dissipation are adequate, a commodity chip can last a long time.
A quick search on Xeon production yields that it goes through a rather rigorous testing. I wouldn't be surprised that server cpu's in a desktop pc works longer. I can't overclock it either, and that probably helps with its lifespan as well. But yeah, the fact that it actually powers on when i click the button and isn't a limiting factor after 10 years is quite something.
Back from my old overclocking days - its heat that kills life. And if you keep that under control (what ages is the heatpaste, replace it ever so often) i very much doubt you'll have any life issues from the cpu itself.
Bearings in fans, caps etc. are also stuff that you need to keep an eye on.
I just replaced a i5-660 thats been powered on since 2010 24/7, heatpaste was fucked so it crashed during heavy loads :)
Old hardware is surprisingly effective. I've been considering a side hustle selling offline AI to local businesses who are privacy-sensitive. Medical, legal, places like that.
At the low end, I'd use old Xeons with gobs of DDR3, install some V100s, run a smaller agent for general chat inquiries, and a frontier model for the deeper stuff, with a router that passes between them depending on the complexity.
The frontier model would perform very slowly, but if it's a deep task the user can submit it in a batch in the evening e.g. "Correlate all of these cases and look for patterns" then receive the output with morning coffee.
Of course, AI helped me work out a plan for this. Haha
The memory controller is integrated into the CPU, so the motherboard chipset is irrelevant. There are some OEM-only v3/v4 parts with dual memory controllers, but the E5-2620 v4 is not one of them.
Doesn't accepting 100% of the MTP draft tokens mean you should just be using the smaller model? Usually the acceptance rate in Qwen36 at least is around 60-70% and the "wrong" tokens are still filled in entirely by the base model, but when you just accept 100% of the draft tokens it seems kind of self defeating unless I'm wrong.
Also I feel like everyone leaves off prompt processing/prefill speeds in these articles. If you are using a very small prompt and asking for mostly generated tokens, sure but I'd love to know the time-to-response of asking for an analysis of an image or a few hundred lines of code.
As far as I know, speculative decoding still verifies that the proposed tokens are what the "big" model would generate, it just uses the guesses to make that process faster. Setting the probability threshold too low then shouldn't affect correctness, just speed (time will be wasted verifying bad guesses).
None of those settings set the speculative decoder to accept 100% of drafted token. I assume you are looking at --draft-p-min 0.0, if so, you are misunderstanding what it does.
It depends on the type of MTP. If you're using two models, draft + full, then arguably yes, the larger model isn't providing much benefit if you really are seeing 100% acceptance rates. There are other forms of speculative decoding that work within the larger model by itself though, eg. Qwen has additional speculative decoding attention heads, so there is no secondary drafting model.
Very intriguing. This might be the use for my e5-2430 V2 X2 server that's been lying around. DDR3 is (relatively) cheap now too. Could fit 192GB of RAM in it and play around for much cheaper than a new GPU.
I've been running various models on a Mac Pro 2013 (8 cores, 32 GB RAM) at about 8 to 10 t/s for months. It's not fast, but it's more than enough for many actual tasks, in particular background tasks. An iMac pro will do just as well I suppose.
The sort of task you don't expect to end immediately. If extracting data from a bunch of PDFs takes 1 hour or the whole night, that doesn't make much difference to me.
It's not fast enough for auto completion and slightly too slow for chat (but bearable IMO).
I have run llama.cpp on an i7-2600 with a 1050. It's too slow for everyday usage but it's not too slow to make it obvious AI is going to be everywhere and in everything. It's too easy to run.
Does this mean my 15 year old Phenom is too old? But it has 16 gb of DDR3 RAM!
Admittedly web browsers and it don't get along that well. Literally the only thing that drags though on my Slackware 15 system, and even then usually only when it gets to around 15 or so open tabs.
Honestly, at this point you're probably looking at a smaller model, for the Gemma series I'd go with Gemma 4 E4B with drafters, but that's just a hunch from using it on my laptop (where I do have a RTX 4060 M and 96gb ram).
So you'd change the invocation slightly here, but a lot of things you can potentially reuse.
That said, the Gemma 4 E4B models have so far in my experience been... not great when it comes to long context, but they are very passable for basic tasks, and even seem surprisingly okay at tool calls.
Have you tested Qwen3.6 35B? Putting aside the capability claims for that model (which I support, but are not my point here), that 35B has smaller active parameter count than the gemma 4 26B, potentially making both prefill and decode faster out of the box, and has MTP heads built in the model and well supported (you may need to make sure you download a quant that didn't strip them off, as some do to preserve space). I would be curious to see your numbers there too. And if you do test this, please go for a clean one and not a fine-tuned one.
i tried the Q4_K_M model form unsloth with your Q4_K_M drafter, but the required memory to load everything is 72GB. odd. otoh i could load Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled.IQ4_XS.gguf and it requires just ~18 GB:
llama_print_timings: sample time = 45.28 ms / 404 runs ( 0.11 ms per token, 8921.67 tokens per second)
llama_print_timings: prompt eval time = 949.42 ms / 51 tokens ( 18.62 ms per token, 53.72 tokens per second)
llama_print_timings: eval time = 24067.08 ms / 400 runs ( 60.17 ms per token, 16.62 tokens per second)
llama_print_timings: total time = 242192.55 ms / 451 tokens
so i wonder why the params used by the quantified qwen model use way less memory than the ones of gemma.
Loading will take some minutes, but at 96 you can squeeze the model in and have some headroom around like ~10 GB, although depending on the Xeon, you may have to downgrade to E4B instead. Should still work thou.
I think one overlooked advantage of older Xeon systems is their availability. Many people can experiment with local AI deployments at a fraction of the cost of building a brand-new setup.
Well, lets get started. I have 4 of those machines, and they are Two dual processor. They all had 32GB of ram, so now I have two with 64GB, and two with zero. They all hand stock K5000s, now how two have two cards. I stripped the uni processors ram and video cards, and put those into the dual procs. They have 256Gb SSDs, and two 1TB disk drives. One machine has 8Gb of VRam across two cards. Dual processors are 8Cx2 and 32 Threads. They can easily play 16 videos at once. For AI, I have not found a model that I can get above 3 tokens a second. Not a one.
That is a very fair point! I just ran a not very scientific benchmark with the system under load, and posted the raw logs in a sibling comment above, but the short answer is that it's hitting 11.94 tokens per second for generation - while it's also being a binary cache and CI build server.
Totally just vibes based, I think it goes up to 20+ tps when it's not under load (and that's me trying to be conservative). For context, reading speed at 250 wpm would be around 5 to 6 tokens per second.
There are ways to trade off compute power for memory bandwidth (like MTP and other speculative decoding approaches). The CPU and GPU would need to be able to share the same cache for this to work. In the Strix Halo case the GPU has a private cache on the GPU die I think, which is the snag.
If you get the inference engine to route the heavy matrix math to the GPU and the speculative drafting to the CPU without choking on latency it's probably gonna be very fast.
Would love to see the benchmarks if someone actually pulls something like that off.
As someone doing this for fun on a windows 11 machine (96gb ram, 5090 24gb) I wonder if I need any flags to keep the model in memory and avoid swapping to ssd?
I use LM studio and qwen3.5 35B - but never figured out if it is swapping or not.
Om am unrelated note, does anyone know a model that can help with this use case:
When you use page up and page down key when reading that blog the first line on the screen is obscured by the floating bar or what ever it is. It is not even needed for reading.
The webpage's layout is just horrible. Scrolling is also
non-default - and thus rather annoying; I had to stop after
two scroll events. Why do people think they need so much
fancy effects or non-standard behaviour, if their alleged
goal is to get information across to other people?
I appreciate the downvotes without any reasoning. It's a fact that newer Intel CPUs have Intel ME which was not in older CPUs and significantly increases attack surface if you are not living in a five eyes state.
In a server, you have to worry about the ME only if you also have an Intel Ethernet interface, which is connected to a potentially hostile network.
If that is not true, the ME cannot be controlled remotely.
The existence of the ME is much more worrisome in laptops, where the ME can be accessed remotely through WiFi. There, to be certain that there is no way for the ME to be accessed remotely you would have to disconnect or cut the internal antennas and use a USB dongle for WiFi.
As five eyes citizen you have at least some rights on paper and you can appeal to your government, but if you are foreigner these guys can go gloves off without any fear of retribution.
Try analyzing Epstein files and posting about it, they'll give you a proper penetration test of all your devices to see what you found out about their ex employee.
Nowadays even EU citizens migrating away from US cloud providers are a "national security issue".
> The argument for speculative decoding is stronger on CPU than on GPU.
Uh. Uuuh.
No?
___
Also
> While a GPU has a massive pool of ultra-fast High-Bandwidth Memory (HBM), a CPU relies on small, lightning-fast “caches” (L1, L2, L3) built directly onto the processor chip.
What purpose does the quoting of "caches" serve there?
Is this AI writing written by that model running on that host?
I ended up getting a modern 26B MoE model (Gemma 4) running at reading speed on an old recycled server with a single Xeon E5-2620 v4 and 128GB of DDR3 RAM (and no GPU). It took a lot of work, but it actually worked out somehow.
I've also linked the quants at the end, but they're not gonna run unless you use the ik_llama-cpp fork I mention, see other posts for more details.
I'm not an ML engineer, so I'm by no means an expert, and the server is busy acting as a Nix cache, but if you have any question, I can try to answer, but best effort.
This poor thing is currently a YouTube watching box.
But ... isnt that a classic use case for SMT? Giving T1 sth. to do while T0 is waiting on DDR(3) and vise-versa?
I also dont understand the explanation of "--cpu-moe". If an expert has ~ 4.0 GiB of Parameters, why does optimizing the sequence of experts minimize cash trashing? With 20 MiB of L3 Cash vs 4.0 GiB of Parameters, it wont cash any noticeable amount of the Parameters, will it?
As mentioned by others, only some Intel Xeon E5-2xxx v4 did support DDR3, and according to Intel, the E5-2620 v4 is not one of them.
Waiting in terms of latency. When the bus is mostly empty and it takes a while to make a round trip it's great to try to find a few extra passengers to put on it. When the buses are all completely full adding the extra riders just makes the bus stop that much more chaotic.
Given knowable runtime hardware usage patterns (huge bursts of memory bandwidth saturation) and a single limited core/thread-shared resource (memory bandwidth), one could optimize for the constraint ahead of runtime.
Because most of the performance optimization levers you have available to pull are (a) trade compute for memory bandwidth (e.g. compression), (b) preload when memory bandwidth is available, (c) optimize the choice of what's in cache when, (d) align to cache size / memory boundaries.
Or tl;dr, try to approximate GPU ISAs at the CPU compiler level. (Which why would anyone but hobbyists, because everyone else just buys pallets of Nvidia/AMD or designs their own ML chips?)
So either you have a v2 instead of a v4 (and run on DDR3 memory), or you have a v4 but with DDR4 memory (not DDR3)
Everything else doesn't work
https://www.intel.com/content/www/us/en/products/sku/92986/i...
Yup that's odd... I've got a Xeon 2680 v4 (14 cores) (amazing bargain of a little beast btw) and it's indeed on DDR4 and I saw all Xeons v4 as supporting DDR4 only.
Full spec (brand/model/mobo type) would have been nice: mine's an HP Z440 workstation repurposed as a server (which I only turn on when I'm working and which I religiously turn off before going to bed).
You say it runs "at reading speed". Have you benchmarked it?
Noted, and agree (it looks like it has also already been clicked, which I dislike). I honestly I need to redo the themes.
> You say it runs "at reading speed". Have you benchmarked it?
At some point a few weeks ago, yes I think so, but I didn't write it down for some reason... so I'll have to find a time when it's not busy and do it again without a noisy system. Right now the system is noisy, but that said doing it like this:
llama-cli --model gemma-4-26B-A4B-it-Q8_0.gguf --model-draft gemma-4-26B-A4B-t-assistant-GGUF/wikitext-2-raw_ik-llama-mtp_drafter-conservative/gemma-4-26B-A4B-it-assistant-Q8_0.gguf --spec-type mtp --draft-max 3 --draft-p-min 0.0 --color -sm graph -smgs -sas -mea 256 --split-mode-f32 --temp 0.7 --cpu-moe -t 8 --flash-attn on --mla-use 3 --merge-up-gate-experts --special --mlock --run-time-repack --spec-autotune --no-kv-offload --parallel 8 --jinja -p "Why is the sky blue?" -n 128
Gives:
So 11.94 tokens per second while it's also playing binary cache and CI builder.When I do it properly, I'll add it to the blog as well!
2010s Javascript, putting down the controller: Ha, no one will ever surpass my high score for wasting programmer time with dependency churn...
2026 Open Source ML: Hold my beer.
A GPU typically processes close to 1000 tokens/s during eval.
It's probably too small for the timings to be taken seriously.
Hyperscalers can perform this evaluation very quickly because evaluation can be significantly parallelized. The layer `i` output of token `j` only requires access to the layer `i-1` output of all previous tokens, so a parallel frontier develops. Token (0,0) [(token, layer)] is processed first, then tokens (0,1) and (1,0) can be processed in parallel, then (0,2), (1,1), and (2,0), and so on.
The maximum parallel width becomes equal to the number of layers in the model. Gemma 4 26B-A4B model discussed in this article evidently has 30 layers, giving a 30-fold speedup if the system were otherwise unconstrained (all layers can be run in parallel, and one full set of layer outputs is completed in the KV pass for each pass of the parallel sweep).
In the specific output above, however, the input prompt is only seven tokens long so there are probably considerable non-amortized spinup effects at play.
The test prompt above was "Why is the sky blue?", so there's the seven tokens. I meant to highlight that because I'd expect processing of a thousand-token input to be faster per token than presented.
As long as performance is useable (apply your own metrics!), pulling it from existing hardware is likely the option with the lower eco footprint.
Also: chances are it'll only be used for this purpose occasionally, and/or for a short while. In that scenario [fabricating new hardware] always has the bigger eco footprint.
If you’ve got something consuming 100 watts average over your 24 hour period, and your electricity costs 20 cents per kWh, you’re already spending almost as much as a Claude subscription.
Just on electricity, this assumes your hardware never fails and you never incur any additional costs.
There’s a big reason why newer more efficient hardware is in demand. Something that’s 10+ years old has drastically worse performance per watt.
Obviously I am not saying to throw away your old hardware as a rule but there is a point where some of this old stuff just isn’t even worth running.
Claude subscription pricing is a broken way to consider footprint.
For a long time, too. Programming languages rarely change much, techniques rarely change, so I should be able to use said model for I hope at least five years; and if at any time they optimize local models to cram even more intelligence into the same amount of VRAM, I can upgrade to that.
I like this path.
I don’t know OpenAI’s infra, but to the extent they are buying GPUs and building data centers with their own money, that sounds like a bad move.
Satya has mismanaged the AI transition in many ways, but one thing he got right is that models are commodities, and the value is in applications that apply them to create user benefit. I agree that any company trying to build a moat with a model is not long for this world.
Paradoxically, the better results we get from general harness of coding agents, the less moat Claude and co. get. It's unbelievably how fast some open models outpaced frontier models of just a few months ago.
Sadly - it's going to be ads. Advertising is going to get in there and enshittify the whole thing because as always, advertising income is too easy and too plentiful for any company to resist.
Right now the models are fairly agnostic, but we are a hair-breadth away from ChatGPT responding with, "the right tool for this job is a circular saw - something like the Milwaulkee M18, which happens to be on sale at Home Depot this weekend."
Enough to validate repurposing an existing workstation with enough RAM, or finding a used high VRAM GPU, or in my case buying a Strix Halo system for home lab and local models.
The future is once again not cloud based, for AI tools.
It makes sense to show some ads and get some money at low volume (like a faraway reader wanting to read a story in your local newspaper) but taking money from regular users directly will pay much more.
Newspapers are happy to cannibalize 99% of their ad revenue with a paywall if that 1% subscribes because that’s how much more money you make from someone paying $10-$20/month vs ads.
But yeah, if people use it as a buying recommendation engine, that’s where the money is on ads/referrals but a lot of AI use has little/no connection to buying intent touchpoints.
LLMs may or may not be able to cover their costs with it. We'll see - I suspect product placement as recommendations will become a thing as it won't take as much GPU to give a "recommendation" on "the best widget for X". I firmly expect it to become enshittified the same way google and amazon search has.
And that's if LLMs don't become commodified.
I think you've misunderstood what good enough means in the context - which is a model capable of completing the tasks assigned to it without having the breadth of full generalization. Your analogy breaks down because of this - we did get 'good enough' spec profiles for different hardware. That thing you're wearing on your wrist won't have the same specifications as the box you use to play games.
> a model capable of completing the tasks assigned to it
The thing is, the "task assigned to it" is changing with improved capabilities. If everyone around you in 2036 is using general AI to do amazing stuff, you will probably have little interest in vibe coding slop like it's 2026.
Only if you give in to fads and FOMO.
The core tasks people need change at a much smaller pace.
That would be the dream... no fucking Electron! No lockdown modules.
That's correct. The problem is they have smart people, tons of money, and several years to figure that out, and the best thing they can come up is a coding agent.
I run my word processing software on my apple 2 (a total joke of a computer) instead of running it on the WANG.
I run my book keeping software on visicalc instead of the IBM.
I run my simulation software on my IBM PC (I even paid for the 8087!) instead of the VAX.
Moore's law has, at least so far, allowed the pioneers with toy computers to grow their toys big enough to solve "big boy" problems after some time has allowed the toy computers to be faster and the pioneers have scaled their crappy home-grown solution to solve their 60% of the problem that was originally solved by some enormous complex system.
Eventually the toy infrastructure gets expensive and solves 90-120% of the "big iron" problem space, but it also grows to cost as much as the big iron solution, but then a new generation of toy software and toy systems emerges to disrupt the "big iron" systems.
See also http://www.catb.org/jargon/html/W/wheel-of-reincarnation.htm...
If a vendor can SaaS a solution, then enterprise is generally happy (they don't want to have to hire folks for maintenance), and that completely locks out any ability to run locally.
Between enterprise's ambivalence and the obvious financial incentive to vendors, you get SaaS-only products.
Make the local AI competent enough to do good image generation and editing, realtime voice and music generation, handle agentic tasks with a framework like Hermes, and you can take your AI places to do tasks in contexts that are inaccessible to or inappropriate for cloud.
Frontier big platform models will be the best, but there's a level of "good enough" for local uses that we're already seeing flourish, and "good enough" for the average joe is almost here.
People -- WANT -- this technology on their home devices and (apparently?) the providers of this tech don't seem to be running a profit so they probably don't want the maintenance tail on their side either.
I think it's a bit different. Inevitable that this becomes a household-run thing? Not likely.
The primary feature of "AI" is to process information and reason with a natural language interface at speed, the primary feature of AI bigboys is to provide the machinery that runs the "models".
See the difference?
But my downtimes are a bit self-inflicted: changing ISPs which I can personally workaround but harder for a blog where one expects uptime.
EDIT: I stand corrected, 200W is apparently way too high of an estimate. I used to run a bunch of old Xeon servers and they slurped watts like crazy, but I can't remember which ones exactly those were.
There's a lot of budget hosting built around chips like these, and they're suprisingly power efficient.
[1] https://www.intel.com/content/www/us/en/products/sku/92986/i...
You likely need to replace the flow-through server chassis system with an active "normal" cooler to achieve a bit of silence.
85W might be about right. My old server CPU is in the same ballpark and compiling kernels it reached about 90w in power usage. If you want to keep it running: idle is not very low power unless you have one of the "low power" L versions, keep that in mind.
Well, you can use it for lots of other things as well.
Compared to the cloud you can probably save up to buy a new server every month. And don't underestimate the gains of having something to experiment on and play with.
Here's my setup. You may want to figure out what the best optimizations are for your specific CPU like AVX2 because mine didn't have most of them. I did try MTP briefly but I wasn't getting performance improvements. You could play around with the batch sizes for cache or context or go even lower for Q2 and don't overcommit on threads either, but I would suggest either defaults or trying out llama-bench. This isn't by any means the best I assume but it worked decently for me and I sometimes swap out Gemma for Qwen. You could also lower q8_0 to q4_0 for more context but it could hurt quality some say, altough I have noticed it too on some models.
# Building
cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=ON -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DGGML_OPENMP=ON
# Running
export OPENBLAS_NUM_THREADS=4
export OMP_NUM_THREADS=4
OPENBLAS_NUM_THREADS=4 OMP_NUM_THREADS=4 \
llama.cpp/build/bin/llama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_XL --temp 1.0 --top-p 0.95 --top-k 64 --min-p 0.00 --jinja --host 0.0.0.0 --port 8080 --cache-type-k q8_0 --cache-type-v q8_0 --threads 4 --threads-batch 4 --ctx-size 8192 -n 8192 --batch-size 2048 --ubatch-size 512 --no-mmap --mlock --chat-template-kwargs '{"enable_thinking":false}' --no-mmproj -np 1 -fa 1
An impressive effort, and better than I would have thought possible on this hardware -- but still pretty far short of what one needs for an satisfactory interactive session.
Should the market react to the memory shortage, the progress of the Apple silicon continue at the same pace, and what we’ll be able to run locally in 6 years will be very exciting. or frightening.
Also I don’t know what this means for the valuation of the AI companies. I remember asking about this very idea to one of their employees at an event and instead of answering he bailed out to grab a cocktail.
- There is no "moat" (lasting, easy-to-defend technological edge) in AI model businesses. There are just short-term advantages.
- An AI business is a capital-intensive business, just like old factories. Data centers are expensive, models are energy-hungry, and the hardware inside must be replaced every 3–4 years.
- Smaller, specialized models eat margins from below. Transcription, voice, or image detection do not need large models.
There is no reason to expect high margins like you can in traditional software business. Benefits of AI go mostly to consumers.
edit: There is potential for economies of scale. Few megacorps can strive for cost advantage when they achieve scale (Microsoft, Google, Amazon and Meta)
It does seem like the structural characteristics we’ve observed so far suggest there is a kind of flywheel from short-term to long-term advantage due to the capital requirements at various levels.
If you’re Nvidia, making the best GPUs today, the expanding wavefront of demand is consuming them with volume and margins to give you a huge edge in building out the best next generation of GPUs. Similar to how the mobile wave gave TSMC sustained advantage for about a decade now.
I’m guessing this is also what we’re seeing as Anthropic and OpenAI swap spots in the token-vendor market.
If you get a not-quite-the-best gaming GPU like a 5080, you can run local models that are better than the state of the art from early 2025. Depending on what you want to do, you might have to switch models. The one size fits all huge models are still a data center thing.
Probably nothing. Most users have no idea what an LLM is or how it runs. Anecdotally speaking, I see many LLM users default to whatever their day job provides to them. And even slightly more sophisticated users seem ok with paying for their openai or anthropic subscriptions.
Maybe we will see a small but dedicated group of open weight model users who prefer local llm, but everybody else will just consume from the big providers? The scenario might look something like OS choices today - a small, committed group of Linux users vs the vast majority of other users running Windows, MacOS, or Chrome?
A new game is a totally new world with everything created from scratch. A creation. A model, on the other hand, is a reinterpretation machine for hundreds of years of human creations, but not a creation in itself, more like a discovery.
You would think that by now we would have a much better Bitcoin that's taking over the payment networks of the world but what we actually got is a shitload of shitcoin.
Which makes sense I suppose.
Even then, if a commodity chip isn't pushed full tilt at all times, and assuming that the venting and dissipation are adequate, a commodity chip can last a long time.
Bearings in fans, caps etc. are also stuff that you need to keep an eye on.
I just replaced a i5-660 thats been powered on since 2010 24/7, heatpaste was fucked so it crashed during heavy loads :)
[0]: https://github.com/ggml-org/llama.cpp/pull/23398
High-Performance AI on a Budget: Optimizing llama.cpp for Qwen3.5 Inference on a Dual-GPU HP Z440
https://news.ycombinator.com/item?id=47320244
At the low end, I'd use old Xeons with gobs of DDR3, install some V100s, run a smaller agent for general chat inquiries, and a frontier model for the deeper stuff, with a router that passes between them depending on the complexity.
The frontier model would perform very slowly, but if it's a deep task the user can submit it in a batch in the evening e.g. "Correlate all of these cases and look for patterns" then receive the output with morning coffee.
Of course, AI helped me work out a plan for this. Haha
What was the net effect of the optimisations? How much faster did it get?
(He has a fully maxed out “last Intel” Mac Pro and laments the lack of replacement).
Also I feel like everyone leaves off prompt processing/prefill speeds in these articles. If you are using a very small prompt and asking for mostly generated tokens, sure but I'd love to know the time-to-response of asking for an analysis of an image or a few hundred lines of code.
Admittedly web browsers and it don't get along that well. Literally the only thing that drags though on my Slackware 15 system, and even then usually only when it gets to around 15 or so open tabs.
https://www.techpowerup.com/cpu-specs/ryzen-7-4800u.c2281
It is way too slow
So you'd change the invocation slightly here, but a lot of things you can potentially reuse.
That said, the Gemma 4 E4B models have so far in my experience been... not great when it comes to long context, but they are very passable for basic tasks, and even seem surprisingly okay at tool calls.
~/ik_llama.cpp[main]$ build/bin/llama-cli --model ~/models/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled.IQ4_XS.gguf --spec-type mtp --draft-max 3 --draft-p-min 0.0 --spec-autotune -cnv --color --jinja --special -smgs -sas -mea 256 --temp 0.7 -t 6 --parallel 6 --cpu-moe --merge-up-gate-experts --flash-attn on --mla-use 3 --mlock --run-time-repack --no-kv-offload . works pretty fast, at about 15 t/s:
llama_print_timings: sample time = 45.28 ms / 404 runs ( 0.11 ms per token, 8921.67 tokens per second) llama_print_timings: prompt eval time = 949.42 ms / 51 tokens ( 18.62 ms per token, 53.72 tokens per second) llama_print_timings: eval time = 24067.08 ms / 400 runs ( 60.17 ms per token, 16.62 tokens per second) llama_print_timings: total time = 242192.55 ms / 451 tokens
so i wonder why the params used by the quantified qwen model use way less memory than the ones of gemma.
Plus many boards also support CXL for RAM expansion over PCI 5!
Source: building a hybrid inference business for regulated industry workloads.
https://pcpartpicker.com/products/motherboard/#s=20028,20029...
This works.
I'd love if anyone knows how I might fare with an old Dell R710 with 2 x Xeon 5600 (12 cores total) and 96Gb of DDR3.
Totally just vibes based, I think it goes up to 20+ tps when it's not under load (and that's me trying to be conservative). For context, reading speed at 250 wpm would be around 5 to 6 tokens per second.
Would love to see the benchmarks if someone actually pulls something like that off.
I use LM studio and qwen3.5 35B - but never figured out if it is swapping or not.
Om am unrelated note, does anyone know a model that can help with this use case:
https://news.ycombinator.com/item?id=48301635
numactl --membind=1
so it is constrained to one of the memory sticks which speeds up token generation a little.
If that is not true, the ME cannot be controlled remotely.
The existence of the ME is much more worrisome in laptops, where the ME can be accessed remotely through WiFi. There, to be certain that there is no way for the ME to be accessed remotely you would have to disconnect or cut the internal antennas and use a USB dongle for WiFi.
As for the second part, I am not sure about how living in a five eyes state would mitigate it. What do you mean by that?
Try analyzing Epstein files and posting about it, they'll give you a proper penetration test of all your devices to see what you found out about their ex employee.
Nowadays even EU citizens migrating away from US cloud providers are a "national security issue".
Uh. Uuuh.
No?
___
Also
> While a GPU has a massive pool of ultra-fast High-Bandwidth Memory (HBM), a CPU relies on small, lightning-fast “caches” (L1, L2, L3) built directly onto the processor chip.
What purpose does the quoting of "caches" serve there? Is this AI writing written by that model running on that host?