I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me to jump into 2 sparks, running the native quant of DSv4 Flash.
Now at 40-50tok/s generation and ~2000 tok/s prefill with a model that I've seen reason through race conditions and be able to trivially pull off any straight-forward coding task, and remain coherent at 500k context. With a preview checkpoint of the weights!
I'm excited for the future of local LLMs. There is some buy-in but apparently not an extreme amount to get access to models that can stand in the for the giants on all but the most challenging and/or hands-off coding tasks.
I suspect DwarfStar could probably squeeze more performance out of the single spark, maybe up closer to 20tok/s.
Moving to 2 sparks meant switching to vLLM with 2-way tensor parallelism and working multi-token prediction. The parallelism and MTP on top of better tuned kernels[1] gave an extremely nice boost! I was quite pleased. I've seen bursts up to 60tok/s at ~150k context - sometimes the MTP seems to really kick in (i.e. high acceptance rate on its tokens)
Currently running a custom vLLM build put together by some folks on the Nvidia forums[2], which speaks to how early support for the model is.
Personally, I've tried to squeeze more tok/s for a single DGX Spark deployment and DeepSeek V4 Flash but only got marginal improvements. There's work to do on fusing kernels and other optimizations that are already on antirez's roadmap so it is not worth duplicating efforts.
I've had positive experiences running GLM 4.7 via vLLM, tool calling works well and the inference is fast. Do you run DeepSeek V4 Flash on vLLM?
I’ve been considering a move to local llm setup, having been underwhelmed coat vs value of various online offerings. But at the same time worried anything I get will be obsolete in a couple months. And I don’t want to have to babysit it. I really want some agents managing and creating side hustles for me and have some other things. I’m technical-have written my own harness and use gh copilot and grok daily and have a hosted openwebui+openrouter thing. I’m also torn between a 128g MacBook Pro or a framework, or spark or similar and lightweight laptop to access. Would love advice anyone has for (or against) going local. I have asked ai but have analysis paralysis as 5k would be a big investment for me so I want to make right choices
Well, if you are making side-hustle money now using online models that, critically, you could also run at home, then it sounds like it’s just a matter of numbers. Oh and, unless you spend a lot more than 5k, your local model will still be slower than the online model. What’s your estimated ROI?
Assuming that’s not true based on your phrasing, you’d be shooting yourself in the foot. Start using online models with the same quant at least benchmark as what you could run at home. Prepare for the at home model to be slower.
You probably want to try renting some time on a dedicated box with roughly the specs you’re considering and running the open models for a bit to see if you would actually use them before dropping a lot on local hardware. A 128 gig MacBook Pro isn’t going to get you an amazing model, and certainly not amazing speed. GLM 5.2 wants something like 350+ gigs at fp4 iirc.
I ran glm 5.2 on rented 8x h200 it could only do 2x concurrency at a cost of $40 an hour. It felt great but dang I wish it was cheaper... It needs 750 at fp8
> You probably want to try renting some time on a dedicated box with roughly the specs you’re considering and running the open models
You don't even need to go that far. For example, with Exoscale Dedicated Inference[1] you just point it at the Hugging Face for the model and quantisation you want to test and it automagically spits out an OpenAI-compatible API endpoint.
> I think they're just suggesting renting as a way to test
Well, yes, I understood that.
Which is why I started with the words "You don't even need to go that far.".
To re-phrase what I said in clearer terms:
Instead of renting an instance, then messing around with configuring Linux and whatever via SSH or Ansible or whatever. Just point a Hugging Face link at this service and get a ready-to-go API back.
Ultimately the guy wants his own hardware. So why waste more time than you need to messing around with someone else's VM.
Have you tried llama.cpp with unsloth and models suited to it? GLM flash? It seemed to allow more models to be tried soon after they are released. Haven’t tried for long term deployment though, that’s the next step.
Highy anecdotal: I have tried various self-hosted models using both vllm and llama.cpp. I am in a situation where I have access to large amount of memory (~320 GB).
While experimenting with quantization I found that there is a non-trivial tradeoff between quality and memory footprint. Overall my experience follows the reported pattern of "2-bit is mwah, 4-bit half decent and 6-bit required for programming. Still, although MiniMax-m2.7 is useable with the 6-bit quantizations that unsloth provides, it felt like such a breath of fresh air when I used the reference full-size model.
I find it difficult to say why. I had mostly the same setup as before (parsing had to be slightly adjusted in Zed). Aside from not experiencing the thinking loops (where minimax would get stuck generating the same sentences over and over) there is little evidence of any real improvement (although the average thinking time felt shorter).
I would recommend against very low quantizations of GLM 5.0/5.1/5.2 or Kimi 2.5/2.6. Smaller models were more reliable, and therefore more useful.
I have tried llama-cpp, vllm is nicer (ray, handles queueing, doesn't have the cache invalidation bug for qwen/gemma models) and unsloth has toxic employees in their discord.
I've run 2 qwen/gemma @8bit with full context window side-by-side. Right now I have 4 models on my spark (qwen36moe, embedding, reranker, qwen3-1.7B) to support my markdown kb tool.
The setup is not as capable, but still good and gets better with models/algos. To me, it's more about the freedom to tinker, freedom from token bill anxiety, and potential right to compute should the government/oligarchy decides it gets to decide who can access which models.
Author here. Quick context the post doesn't quite spell out:
The tool_choice="auto" failure on Qwen3-Next isn't a parser issue — the model
reasons inside <think>, decides, and never emits the tool call. No error, just
empty tool_calls. The fix was swapping the backbone from Thinking to Instruct,
not tuning any parser flag.
The "load the bigger model first, size the smaller against actual residency"
playbook generalizes to anything with shared CUDA framework overhead. The ~5 GiB
framework floor shows up even at small gpu_memory_utilization values — plan
against actuals, not targets.
Can you try and tune your Claude or whatever LLM you're using for your text to phrase things in plain English. Way less use of antithesis, at least. You can probably find a skill for it, if not get an LLM to write your own.
Now at 40-50tok/s generation and ~2000 tok/s prefill with a model that I've seen reason through race conditions and be able to trivially pull off any straight-forward coding task, and remain coherent at 500k context. With a preview checkpoint of the weights!
I'm excited for the future of local LLMs. There is some buy-in but apparently not an extreme amount to get access to models that can stand in the for the giants on all but the most challenging and/or hands-off coding tasks.
[1]: https://github.com/antirez/ds4
Not clear how you went from ~11-14 to ~40-50 tok/s. Is it by running the quant native model and adding a second spark?
Cheers
Moving to 2 sparks meant switching to vLLM with 2-way tensor parallelism and working multi-token prediction. The parallelism and MTP on top of better tuned kernels[1] gave an extremely nice boost! I was quite pleased. I've seen bursts up to 60tok/s at ~150k context - sometimes the MTP seems to really kick in (i.e. high acceptance rate on its tokens)
Currently running a custom vLLM build put together by some folks on the Nvidia forums[2], which speaks to how early support for the model is.
[1]: https://github.com/lukealonso/b12x
[2]: https://forums.developer.nvidia.com/t/372268
I've had positive experiences running GLM 4.7 via vLLM, tool calling works well and the inference is fast. Do you run DeepSeek V4 Flash on vLLM?
Assuming that’s not true based on your phrasing, you’d be shooting yourself in the foot. Start using online models with the same quant at least benchmark as what you could run at home. Prepare for the at home model to be slower.
You don't even need to go that far. For example, with Exoscale Dedicated Inference[1] you just point it at the Hugging Face for the model and quantisation you want to test and it automagically spits out an OpenAI-compatible API endpoint.
[1] https://www.exoscale.com/ai-cloud-infrastructure/dedicated-i...
(I have no relationship with Exoscale, this particular product just crossed my radar recently)
Well, yes, I understood that.
Which is why I started with the words "You don't even need to go that far.".
To re-phrase what I said in clearer terms:
Instead of renting an instance, then messing around with configuring Linux and whatever via SSH or Ansible or whatever. Just point a Hugging Face link at this service and get a ready-to-go API back.
Ultimately the guy wants his own hardware. So why waste more time than you need to messing around with someone else's VM.
While experimenting with quantization I found that there is a non-trivial tradeoff between quality and memory footprint. Overall my experience follows the reported pattern of "2-bit is mwah, 4-bit half decent and 6-bit required for programming. Still, although MiniMax-m2.7 is useable with the 6-bit quantizations that unsloth provides, it felt like such a breath of fresh air when I used the reference full-size model.
I find it difficult to say why. I had mostly the same setup as before (parsing had to be slightly adjusted in Zed). Aside from not experiencing the thinking loops (where minimax would get stuck generating the same sentences over and over) there is little evidence of any real improvement (although the average thinking time felt shorter).
I would recommend against very low quantizations of GLM 5.0/5.1/5.2 or Kimi 2.5/2.6. Smaller models were more reliable, and therefore more useful.
I've run 2 qwen/gemma @8bit with full context window side-by-side. Right now I have 4 models on my spark (qwen36moe, embedding, reranker, qwen3-1.7B) to support my markdown kb tool.
The setup is not as capable, but still good and gets better with models/algos. To me, it's more about the freedom to tinker, freedom from token bill anxiety, and potential right to compute should the government/oligarchy decides it gets to decide who can access which models.
Edit: 3.6 not 3.7!
Someone's optimistic
The tool_choice="auto" failure on Qwen3-Next isn't a parser issue — the model reasons inside <think>, decides, and never emits the tool call. No error, just empty tool_calls. The fix was swapping the backbone from Thinking to Instruct, not tuning any parser flag.
The "load the bigger model first, size the smaller against actual residency" playbook generalizes to anything with shared CUDA framework overhead. The ~5 GiB framework floor shows up even at small gpu_memory_utilization values — plan against actuals, not targets.