This is more exciting than k3, IMO. Dsv4 models are extremely cheap to serve. Improving their capabilities has lots of downstream effects, as it becomes "good enough" for more and more tasks.
DS was serving the pro version at extremely low prices for a long time, and they've had integrations with opencode & other providers, so they likely gathered a lot of data from real developers doing real tasks (on openrouter they were labeled as such). Now they can use those live scenarios to further post-train their models and improve them further.
Can't wait to see if distilling k3 into dsv4 brings additional improvements. Anyway, having fast cheap models getting better is great for the community. Especially since these don't "go away" on a provider's whim. Whatever capabilities they get, can be used "forever" going forward. And, at least flash can be ran "at home" with <10k in hardware, which isn't really possible / feasible with glm/k3 larger models.
Qwen/Alibaba have stopped doing open weights releases for a while. No grudge or anything, I'm certainly not going to look at a gift horse in the mouth, but both DeepSeek and Moonshot have been very consistent with open weights as well as sharing actually detailed research.
In terms of open research, China has absolutely overtaken the US.
Qwen has a pinned tweet stating that 3.8 will be released as open weights soon. I guess it remains to be seen, though, if they’ll do the smaller model sizes or only the big 2.4T one.
What are your goalposts? Depending on your requirements, there have been many moments of usable local AI. More recent ones were gpt-oss 120b and Qwen 3.6 27b.
I've been driving flash model for 90% of my tasks. It's better than pro (for unknown reasons), very cheap and fast.
I try to keep changes under 1000 lines and drive architectural decisions myself, barely notice any difference compared to frontier models. The rest 10% is to spot bugs, security problems and to investigate better architecture, which flash can also do pretty well, I just cross check it.
Faster iterations are way better for me, I hate waiting for 5-10 minutes on small changes. I tried to use recent versions of Kimi and GLM, but they use too much thinking for no reason and are pretty slow because of it. I also often feed a lot of data to it, without worrying about hitting the limits: dependencies (to find bottlenecks in them), logs, performance dumps and so on.
Also, it will never complain about security guards, I've been using it to reverse engineer binaries.
This is my experience also: DeepSeek v4 flash is good enough for most of my work and I like the fast response times. I buy tokens mostly from FireWorks.ai in the US, but I also prepaid for a large chunk of tokens directoy with DeepSeek.
I use OpenCode mostly (uses fewer tokens than Claude Code) and I am looking forward to the release of DeepSeek’s own coding harness.
> Also, it will never complain about security guards, I've been using it to reverse engineer binaries.
Maybe I'm using too weak language in my prompts, but none of the OpenAI models I've used via codex has refused to reverse engineer binaries, is it supposed to? I'm sitting right now reverse-engineering a 3rd party firmware together with Codex and haven't hit a single guardrail. Meanwhile, I see people complaining about it rejecting non-security related prompts, are things so individual on the platforms right now or what's going on?
How so? The code and reasoning patterns certainly match, so does the intelligence level. I probably wouldn't know if they're secretly serving Luna instead of Sol, but it's definitely an OpenAI model.
It's good but (at least on openrouter) it's got an annoyingly tight output token limit. So if it does get stuck in a reasoning pit, it won't work its way out of it in time.
I use deepseek for a lot of my personal day-to-day agent needs, and I will simply put this here and let this speak for itself, last 30 days:
- Cost: $4.55USD
- API requests: 3,467
- Tokens: 323,183,886
And as an engineer who leads a small team, I have very high standards for quality, and these carry across to my personal projects where I use deepseek. It has not disappointed at all for coding or review tasks.
For everything else, use another model.
I'm aware that almost all of this is prompt voodoo, and there's no guarantee for the review to find anything or everything, but making it a dedicated skill and thoroughly observing the output thinking, tool calls, and result, lets me adjust these over time and fill the weak spots with even more prompting.
I use this skill by simply telling the agent something like "Review the changes on the current branch against origin/main, take special care with backwards-incompatible changes to the public API" or something like that.
I use `pi` (pi.dev) with a subagents extension, so that I can ask the agent to invoke a subagent to do the review, on work that the agent did.
For models, I use the highest possible reasoning on whatever model I feel like makes sense, usually this is GPT-5.5 or deepseek flash/pro, depending on the confidentiality of the codebase, on the highest reasoning always (for reviews).
This is why my review skill mandates a very strict output contract. I need the output to be very easy to parse, and the output contract I've specified there does that.
In general I let <whatever the latest model of OpenAI's ChatGPT is> author and review SKILL.md and similar large prompts, usually with a ruleset like this, which is a 1600 line research artifact from a long GPT 5.5 "Pro" research session on prompt engineering: https://gist.github.com/lionkor/71498794d0a7d72173fc58766f25...
Does the review catch all issues? Not at all. Does it catch, usually more than one, important issue, across large changesets? Absolutely, and that's the point! :)
Feel free to ask me any questions, I'm also happy to share more about my setup via email or add you or anyone else to my private repos with more of these.
It hallucinates plenty, about the same as Codex models and all other LLMs! I review all code it writes, thoroughly, check the test coverage, write tests myself, have other models/chats cross-check the work with a review skill (which I've shared in another comment in this thread).
Usually unreviewed code only gets committed if I really don't care, like for one-off scripts, which I sandbox with github.com/lionkor/sbh or run as an unprivileged user.
Aside from what squidbeak already said, I have to say the quality varies a lot, and goes above the "good" threshold enough times that using these models is worth the time and effort. Compared to something like GPT-5.4 to GPT-5.6 Codex models, they cost 50-100x less (fifty to one hundred TIMES less), and can do most annoying or well-specified tasks just as well.
The main difference is that deepseek is bad at prose, and Codex models are much more eager to use tools provided to them (which is usually fine, they often make like 3 todos via tool calls for a simple one step task which is silly though), and deepseek in general benefits from good instructions more than GPT models maybe do.
Deepseek becomes much better if you give it tools for asking clarifying questions, doing self-review with subagents, and so on. The more tools it uses, the better the signal-to-noise ratio in its context, and the more consistent the output.
You must still review 100% of the code as if it's trying to sell you insurance.
An experienced software engineer (read his profile) praises the value he's found in Deepseek, and gives some real data showing how affordable that value is.
Then you, dakolli - out of generosity and minute-to-minute devotion to enlightenment - sacrifice time from your busy day to sit down (though perhaps that's been painful lately?) or stand up with your phone - and offer a profound, deeply thought-out counterpoint in the following form (and I'll paraphrase):
His Bio: c++, rust and c#, software engineer located in Germany, currently working in the electrical power industry (high- to ultra high voltage testing- and measurement equipment).
The stories I could tell you about people who would have that exact bio. I work adjacent to this industry, I promise you there are thousands of people with these exact credentials who are extremely bad engineers. Also, I'm not sure why having something in your bio makes you qualified.. You llm people have completely fried your brains.
Sorry, I use these tools every single day, a lot, I'm not speaking from a lack of exposure, but these llms are terrible and all they do is add massive amounts of slop to your code. Yes they're useful for when you want to be lazy, or you're doing something not serious, or for the people who want to be lazy all the time (most llm proponents). Have fun making your competency tied to the quality and quantity of tokens you have access to slopper.
I really can't begin to describe how stupid I think you are for thinking that having that statement in your bio makes you a qualified engineer.
Essentially I'm running everything on flash now inside pi. With the correct set of MCP servers, context reducer tooling and skills it can implement any task I throw at it. Some sessions take 30+ turns, but it's fast and cheap; all this in an hour, with ~$0.5 cost.
(TBH though, in my multi-subagent workflow I do use other, more expensive models for planning, reviewing, oracle-ing)
I haven't used our slow opus subscription for weeks.
(Also set up an OpenWebUi self-hosted chat that works from my phone, has some mcp and skills. fully replaced perplexity. Monthly cost ~$18 for hosting and subscriptions)
I want to second this, I use the same setup (pi + deepseek, with lots of custom tools for tracking TODOs, doing things with less tokens, etc, and with a SOTA model for very difficult tasks) and it's all I need it to be.
Recommendation? No. Just go with the passive-aggressive advice "let pi build it for you". :-)
To be more constructive, what I did (as an experiencd SWE but a complete noob to agentic coding): went to pi.dev's extension marketplace and looked into all the new shiny stuff. Subagents, mcps, context and memory optimizers, skills. Using the most popular ones (not necessarily the best ones)
It was like 15years ago learning the new mindset of vim (and spending a ton of time to customize it to my workflow). My understanding is that Claude and opencode doesn't give you this flexibility.
Learning all these stuff drove me to also set up openwebui, and it was such a successful private project that I implemented it at work (with jira/confluence/bazel query access) and management said "we need this by tomorrow".
I believe the models matter not that much anymore. The "harness" does. (unless you just want to vibe code. Thebn, throw crap at fable and call it a day)
omp is not really pi. Its a very distant fork of it. And its philosophy is completely different, it includes everything you need except workflow skills. Personally, I think it is the best agent coding tool by far, and its what I settled on after trying 6+ different tools.
Last night it single handedly implemented a feature after a grilling session, and came back with the red-yellow-green risk assessment points that I mostly saw with anthropic models. I had to check if I'm using the right model, but it was DS4Flash.
Possibly depending on who's routing it, since Openrouter seems to have added a separate slug for this, and they don't have an alias router for the DS models yet, so it shouldn't have swapped over to it in case you use them for routing.
If you are directly using it through Deepseek, they do mention that the flash slug has migrated over
You build it as an extension or create your own fork/copy of pi and add it. This basically goes for most things in pi, except the most basic stuff. It's basically meant for people who think "I'll just add that myself" rather than expecting it to be there out of the box or finding other's solution to it, for better or worse.
If the benchmarks are real and reflect actual use, then this is an insane model. This 300B model outperforms the previous DS4 Pro preview model (1.8T params), and it looks like it outperforms GPT 5.6 Luna too. And it's still cheaper than Luna, even with the price decrease.
For reference, that puts it on the third spot behind GPT 5.5 and Fable 5. For some reason GPT 5.6 Sol is not showing in the leaderboard. If it did, then DS4 Flash was number four.
The thing is, even if Luna is better in DeepSWE and has the 80% discount. DeepSeek is still cheaper.
--------------
DeepSeek V4 | Flash GPT-5.6 Luna (New)
--------------
Input (Cache Hit) $0.0028 $0.02
Input (Cache Miss) $0.14 $0.20
Output $0.28 $1.20
--------------
Both Luna and Flash are heavy on the reasoning > output. And the cache hitrate + prices also matter.
Reality is, you can not go wrong with Luna or Flash at those prices. And remember, DeepSeek V4 Pro is still in the rafters, what is ironically closer to Luna's new price.
Is anyone using DSv4 for their agents that is not related to writing code? Curious about use cases specially for someone using gpt-5.4 mini for classification, categorization, etc
Objectively speaking, the 2 bit quant from antirez has very low accuracy. Meanwhile, his 4 bit quant does have decent accuracy, but is a bit pointless by being bigger than the full precision MXFP4 quant. Anyway, they all work fine in practice.
Very promising. So it will both keep the speed and reduced price, yet exceed performance of the quite sufficient deepseek-v4-pro?
Should be extending the lead in intelligence/cost index, as deepseek-v4-flash already were the most price efficient model, which now becomes even better. Although, in the deepseek APIs, the cost is leaking all information about codebases to China.
Looks like they will release an updated version of deepseek-v4-pro soon, which most likely will beat kimi k3 at both intelligence and cost (judging by the vast improvements to dsv4 flash)
it does! I'm always amazed when I use DSV4 flash for coding or server checks and after an hour of working with it my (pure api call) balance is about 30 cents
DeepSeek is great for tasks and software I already know well. Even if it gets something wrong, I can usually verify it myself. But when I'm working with a programming language I'm not familiar with, I prefer using Codex or Claude.
For both US and China models - what standard security checks and QaQc are you all doing? We're running small gamuts to test for unsolicited jailbreaks (model jailbreaks you) and incorrect records (Fake Accuracy - as Easter Egg or common thread) meaning falsified logic or information cooked in by the developers, rather than the training data speaking for itself
I was just using it when it landed. It started reasoning more extensively from nowhere and precision went up a lot. It also changed its prose style for the better. Looking forward to weights.
I wouldn't doubt GPT 4.6 Luna being in the top left quadrant's center on the Cost per Intelligence Index is not concerning for Liang Wenfeng. You have to remember DeepSeek v4 flash even though a bit cheaper, does not have vision abilities, which is a big draw for agentic tasks.
I admire DeepSeek's openness, but even they have been raising prices after their discounts.
They haven't raised prices though the plan was to increase it with peak hour usage for V4 Pro GA release, they didn't do that, so no price increases there I believe.
As for vision yeah it sucks but Luna is also 2x input and 1.5x output for 1M context...
That's around 0.4 in/1.8 out
DSv4 is wayyy cheaper.
And it's open now you have Luna at home if you have a decent set of GPUs you can run this on 2Sparks or one very expensive Mac or just like 6-8 5090s..
I know most folks can't afford it I am working on making it viable to rent shared hosting the biggest issue is data leak and prompt injection attacks with shared hosting. (Since the server owner connects to your main system via the coding agent)
I guess using a ZDR provider is good enough for now.
> I wouldn't doubt GPT 4.6 Luna being in the top left quadrant's center on the Cost per Intelligence Index is not concerning for Liang Wenfeng.
The leaked interview has him saying it doesn't matter... as much as open source doesn't matter. There's enough in it for everyone right now and they aren't after everything.
Perspective: DeepSeek doesn't have enough infrastructure to serve their target customers already.
Can you post the link to the leaked interview? From what I understand he has been pretty tight lipped for a guy who has a larger stake worth more in his company than Dario does in Anthropic.
Note: if you are having success with a model, then please post what you are using it for. Writing HTML/CSS is very different from writing Rust/C++ or doing maths.
I do Python, Go and Rust. Go works the best, I would say Python is worse than Go.
Rust works perfectly fine, but when I use Rust, I usually pay closer attention to performance, so I have to guide it a bit, to improve cache locality, use simd, avoid unnecessary allocations and so on. Terra has the same issues with Rust. If you always prompt models to achieve the best performance, the code is usually no the one that I want, they optimize unnecessary/cold parts or blindly optimize stuff where compiler takes care of the optimizations already.
See my other comment for more information on how I work with it. In short, keep the changes under 1k lines, context under 120k (ask it to use subagents), drive the architecture yourself.
I've been having better-than-most performance using qwen3.6 to write python/flask.
I tried using it to generate some rust code yesterday, and it generated much code but only ever came within 1 error of a testable build. The 4th or 5th full rewrite is sitting in the buffer right now.
I'm currently looking to up my game with Bottlecap AI's return of qwen3.6, 'thinking cap'.
Alleged to be twice as fast and superior at coding over extended sessions (vs. 3.6).
Oh my goodness what an update. I need these weights. It's an incredible model for the size. The improved tool calling etc. should be able to make my harness way simpler. This runs at mega-speed on prosumer hardware (2x RTX Pro 6000).
Pairing non-APIs with APIs tend to be a hassle, and risky, as usually that breaks the ToC. I guess easiest for you to test if it's worth using the OpenAI API, is to manually copy-paste responses between wherever you run V4-Flash and the ChatGPT UI. What I've done in the past is basically .zip up the entire project directory, ignoring files from .gitignore, then asking ChatGPT Pro to inspect that and come up with a plan, then you paste that to where you have V4-Flash. Basically how we did "vibe pair programming" before the TUI agent harnesses appeared in the ecosystem :)
reminds me back in the day when we used to do version control with google drive. you generally don't need to do this anymore i don't think. it should be fairly trivial to open the codex cli in the project directory, ask it to do some analysis and planning -> save to md. then opencode with deepseek to read that md and start implementing. consider that some cross-model subagent workflows already exist: i use claude/fable to plan and there are openai plugins to directly spin up codex subagents. this latter approach gives claude/fable more directly control to babysit the codex agents whereas the former is more of a clearcut handoff
That approach won't give you access to the "Pro Mode" nor "Deep Research" though, which if you're paying for a Pro subscription already, "Pro Mode" is probably the top-top when it comes to "extensively researched answers & plans" across the entire ecosystem. That said, I do ask it to produce a .md plan that later is used with agents :)
They're always very understated in their update descriptions. This is actually a HUGE improvement in the model's capabilities rather than just a small tweak.
Where do you guys get deepseek? I'm hearing a lot of good reviews and want to try it with my pi config. from the deeepseek themselves, openrouter, or anywhere else? does it make a difference?
[edit]: whoa it is really fast. will take some time to evaluate quality thou
For hosted APIs, it's a lot cheaper to use their own infra, caching seems a hell of a lot better there compared to OpenRouter, and indeed the tok/s seems higher. They also have peak/off-peak pricing, so if you can hold off with your request, you get a pretty big discount.
Otherwise, if you're trying to run it locally, even really low quantizations like DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix seem to actually not be so dumb compared to smaller models with same quantization, might be worth a try if you're sitting on a lot of RAM/VRAM yet not industry-scale amount :)
Opencode-go gives you $60 worth of DS V4 api usage for $10 per month. Right now I think it's hard to exhaust that when using flash exclusively, and plain API use might even be cheaper! Anyway, for DS usage it's a good deal.
Opencode also have a ZDR (zero data retention) deal with them – if I recall correctly, that's not something you can enable as an individual DeepSeek subscriber.
I really like OpenCode - BUT: there is a loophole the size of Portugal in that ZDR language. All they say is that their providers follow a ZDR policy. I haven’t found anything promising that OpenCode themselves don’t retain Go usage data. Something to be mindful of.
Anecdotal data from my own tests: allow only one provider if you want good cache usage on OR. 80 cents is probably 4x the price you should have paid.
Providers' cache hit stats are available to consult, and only 1-2 of them behave properly if I remember correctly, zero if you request providers that don't store and train on sessions.
Wonder how good the proper version of V4 Pro will be.
I'm still considering pulling the trigger on the annual subscription of Kimi for K3 but it's sometimes slower than I'd like (at least when compared to Anthropic) even on their Vivace plan, and the token limits on the GLM Coding subscription for GLM 5.2 were too easy to hit.
Every time I want to have fun coding something with natural language processing, I use deepseek flash. It's just incredible for the price. I have a fairly popular app with 400 users that uses DeepSeek in the background and it still didn't hit even 50 bucks of usage in a month.
The previous V4 version wasn't called “Preview” by most inference providers. For example, the OpenRouter model slug was `deepseek/deepseek-v4-flash`. So now there will be confusion when someone talks about V4 Flash or when someone offers V4 Flash inference.
It probably would be called 'deepseek-v4-flash-0731' in API
edit: nope, at least deepseek kept "deepseek-v4-flash" and just updated model underneath. I guess preview is no longer worth serving with that release and you'd have to look through inference provider docs to see if they've updated, yeah..
That's what I mean. On DeepSeek it's now just `deepseek-v4-flash`, while OpenRouter calls it `deepseek/deepseek-v4-flash-0731`, so now when someone talks about DeepSeek V4 Flash, like in benchmarks, or other inference providers, which version do they actually mean?
The `-0731` style suffix is worse compared to a proper version bump like V4.1.
What I and my team have been doing in papers is just to refer to the OpenRouter slugs. It upsets reviewers because they will complain it "is not sufficiently clear to a wider audience" but I do agree it's the cleanest approach.
Also goes to prove how much power OpenRouter has actually...
Not important, just a curiosity. I would expect that kind of knowledge would be pretty well burned into the weights, but what do I know about LLMs. Gemma 4 can answer this question well.
Yes that's my point. The old and the new version are different in capabilities, but now when someone talks about DeepSeek V4 Flash (in benchmarks, on inference providers), you don't know which exact version it's about.
Some providers like OpenRouter now call it `deepseek-v4-flash-0731`, but even in places like here on HackerNews people say things like "Sonnet is better than DeepSeek" without specifying a version or a reasoning effort, certainly no one will mention that `-0731` suffix when talking about DeepSeek V4 Flash.
I think it does not matter. Most people who use these types of models are into the IT world, and will know/be informed very fast that there is a difference. People will likely also use the 0731 behind it.
The only issue i see, is 3th party providers that have not yet updated. But that is going to be a short time periode. There is no reason to not update.
Finally have a model with usable intelligence, at a reasonable price.
Can't imagine what Pro GA would look like, considering pro preview has only 1.6t parameters.
Weights are yet to be released (maybe in 24H, Deepseek has track record of releasing same day).
https://github.com/antirez/ds4 is often mentioned for DS4F on Mac, but 64GB is likely not enough to achieve reasonable speeds (official weights should be ~160GB).
The conspiracy theorist in me wants to think that the 80% drop in GPT 5.6 Luna prices today is correlated with this update from DeepSeek. Perhaps OpenAI has already hacked its competitors with it's Mythos-like models and is aware of what competitors are doing and is able to react in advance.
In case people want to run it, it's DeepSeek-V4-Flash-284B-A13B. So it should just barely run on a single B300, and it's small enough that it'll barely run on an M5 Max too.
Probably the same. When the same base model is trained, the weight do not tend to change a lot. GLM 5.0 > 5.1 > 5.2 are the same base model, that just kept being trained. Weights hardly change as a result. Think in the like few percentage points size difference.
I'm running a useful quantization of the previous version of Deepseek-V4-Flash -- quite well but with so much fan noise -- on a MacBook Pro M5 Max with 128 GB.
There's also the 2x spark way, which should be ~8k eur? Someone down the thread reported ~60tps for 2x sparks. That's totally usable for local inference.
You can also do 2x 6kPRO in a workstation, for ~20k.
The memory bandwidth of the 2x RTX Pro 6000 Blackwell setup will be 10x higher, which should have an equivalent effect on the generated tokens per second.
Currently the 3bit (and 2 bit) quant on DGX spark (on one of them) and the M5 Max should just start. Right now. (I'm hoping to get an M5 Max delivered on monday, let's see if it happens this time. It's 2+ months since I ordered now)
The 4 bit quant technically fits (there's a 127 GB version) but ... obviously that's not going to work. It is so close though, surely someone will a way to do it.
Chinese labs rushing to release models this week, because it's inevitable that Washington regulates Chinese models in the next 4 weeks. All the US AI leaders have been taking trips to Washington this week, what do you think they're there for..
Open flash model is competing against OpenAI's 'Sonnet' model at the price of GPT 3, I am really excited about this release, hopefully it holds up in real work as well
IIRC GPT 3 was priced at per 1k tokens, had to check, the biggest GPT 3 model from OpenAI was $0.06/1k, so $60 / 1M. gpt-3.5-turbo was the first model after ChatGPT and that was $2 / 1M. And no caching. So not really in the same ballpark
They're literally comparing the previous version of the same model with the new one. It's based on the same architecture, same pre-trained model, just different post-training. It doesn't get more apples to apples than this.
Ah, my bad. Yeah that makes sense. They do say "The official V4-Flash natively supports the Responses API format and is specifically adapted for Codex.", so at some point someone will make a "same harness" comparison.
It will be fair if they release the harness though. I think now the future will be paired model-harness releases, not just weight dumps.
The performance changes are so big with the right harness that is makes sense to engineer the harness and fine-tune the model to one another from the start.
Open weights means that there is a free market on providing inference using this model. Some of them will offer good data protection. (Well, hopefully)
DS was serving the pro version at extremely low prices for a long time, and they've had integrations with opencode & other providers, so they likely gathered a lot of data from real developers doing real tasks (on openrouter they were labeled as such). Now they can use those live scenarios to further post-train their models and improve them further.
Can't wait to see if distilling k3 into dsv4 brings additional improvements. Anyway, having fast cheap models getting better is great for the community. Especially since these don't "go away" on a provider's whim. Whatever capabilities they get, can be used "forever" going forward. And, at least flash can be ran "at home" with <10k in hardware, which isn't really possible / feasible with glm/k3 larger models.
In terms of open research, China has absolutely overtaken the US.
You can’t treat Qwen3.6 like its fable, but if you prompt precisely and specifically it’s a great executor.
I actually found it refreshing to use more of my brain for once, and actually have to think deeper about what I’m trying to do, and how to build it.
I try to keep changes under 1000 lines and drive architectural decisions myself, barely notice any difference compared to frontier models. The rest 10% is to spot bugs, security problems and to investigate better architecture, which flash can also do pretty well, I just cross check it.
Faster iterations are way better for me, I hate waiting for 5-10 minutes on small changes. I tried to use recent versions of Kimi and GLM, but they use too much thinking for no reason and are pretty slow because of it. I also often feed a lot of data to it, without worrying about hitting the limits: dependencies (to find bottlenecks in them), logs, performance dumps and so on.
Also, it will never complain about security guards, I've been using it to reverse engineer binaries.
I use OpenCode mostly (uses fewer tokens than Claude Code) and I am looking forward to the release of DeepSeek’s own coding harness.
Maybe I'm using too weak language in my prompts, but none of the OpenAI models I've used via codex has refused to reverse engineer binaries, is it supposed to? I'm sitting right now reverse-engineering a 3rd party firmware together with Codex and haven't hit a single guardrail. Meanwhile, I see people complaining about it rejecting non-security related prompts, are things so individual on the platforms right now or what's going on?
It's replaced the Kimi models for me though.
- Cost: $4.55USD
- API requests: 3,467
- Tokens: 323,183,886
And as an engineer who leads a small team, I have very high standards for quality, and these carry across to my personal projects where I use deepseek. It has not disappointed at all for coding or review tasks. For everything else, use another model.
I've had worse experiences doing it because the quality of answer has been quite bad, and I'm wondering if my methods are the reason.
Review is a skill, as in, a SKILL.md with a folder full of references:
- SKILL.md: https://gist.github.com/lionkor/161525be858d1d75db4c13c0f093...
- references/output-contract.md: https://gist.github.com/lionkor/8c68e33becef7a21f8408c7dc119...
- references/review-lenses.md: https://gist.github.com/lionkor/0a8b080fe45306213efddf3ebb75...
- references/review-workflow.md: https://gist.github.com/lionkor/d2d374b133ceb7e3660bd530ee72...
- references/section-rules.md: https://gist.github.com/lionkor/8a9e503adc7fd3697410cf021f27...
I'm aware that almost all of this is prompt voodoo, and there's no guarantee for the review to find anything or everything, but making it a dedicated skill and thoroughly observing the output thinking, tool calls, and result, lets me adjust these over time and fill the weak spots with even more prompting.
I use this skill by simply telling the agent something like "Review the changes on the current branch against origin/main, take special care with backwards-incompatible changes to the public API" or something like that.
I use `pi` (pi.dev) with a subagents extension, so that I can ask the agent to invoke a subagent to do the review, on work that the agent did.
For models, I use the highest possible reasoning on whatever model I feel like makes sense, usually this is GPT-5.5 or deepseek flash/pro, depending on the confidentiality of the codebase, on the highest reasoning always (for reviews).
I've also had success with a review checklist, though it doesn't produce an easy to parse (for humans) output: https://gist.github.com/lionkor/054ac2cf241e0765eee2383f0dba...
This is why my review skill mandates a very strict output contract. I need the output to be very easy to parse, and the output contract I've specified there does that.
In general I let <whatever the latest model of OpenAI's ChatGPT is> author and review SKILL.md and similar large prompts, usually with a ruleset like this, which is a 1600 line research artifact from a long GPT 5.5 "Pro" research session on prompt engineering: https://gist.github.com/lionkor/71498794d0a7d72173fc58766f25...
Does the review catch all issues? Not at all. Does it catch, usually more than one, important issue, across large changesets? Absolutely, and that's the point! :)
Feel free to ask me any questions, I'm also happy to share more about my setup via email or add you or anyone else to my private repos with more of these.
Usually unreviewed code only gets committed if I really don't care, like for one-off scripts, which I sandbox with github.com/lionkor/sbh or run as an unprivileged user.
The main difference is that deepseek is bad at prose, and Codex models are much more eager to use tools provided to them (which is usually fine, they often make like 3 todos via tool calls for a simple one step task which is silly though), and deepseek in general benefits from good instructions more than GPT models maybe do.
Deepseek becomes much better if you give it tools for asking clarifying questions, doing self-review with subagents, and so on. The more tools it uses, the better the signal-to-noise ratio in its context, and the more consistent the output.
You must still review 100% of the code as if it's trying to sell you insurance.
Then you, dakolli - out of generosity and minute-to-minute devotion to enlightenment - sacrifice time from your busy day to sit down (though perhaps that's been painful lately?) or stand up with your phone - and offer a profound, deeply thought-out counterpoint in the following form (and I'll paraphrase):
"Nah mate, it's shit. All LLMs are shit."
The stories I could tell you about people who would have that exact bio. I work adjacent to this industry, I promise you there are thousands of people with these exact credentials who are extremely bad engineers. Also, I'm not sure why having something in your bio makes you qualified.. You llm people have completely fried your brains.
Sorry, I use these tools every single day, a lot, I'm not speaking from a lack of exposure, but these llms are terrible and all they do is add massive amounts of slop to your code. Yes they're useful for when you want to be lazy, or you're doing something not serious, or for the people who want to be lazy all the time (most llm proponents). Have fun making your competency tied to the quality and quantity of tokens you have access to slopper.
I really can't begin to describe how stupid I think you are for thinking that having that statement in your bio makes you a qualified engineer.
(TBH though, in my multi-subagent workflow I do use other, more expensive models for planning, reviewing, oracle-ing)
I haven't used our slow opus subscription for weeks.
(Also set up an OpenWebUi self-hosted chat that works from my phone, has some mcp and skills. fully replaced perplexity. Monthly cost ~$18 for hosting and subscriptions)
To be more constructive, what I did (as an experiencd SWE but a complete noob to agentic coding): went to pi.dev's extension marketplace and looked into all the new shiny stuff. Subagents, mcps, context and memory optimizers, skills. Using the most popular ones (not necessarily the best ones)
It was like 15years ago learning the new mindset of vim (and spending a ton of time to customize it to my workflow). My understanding is that Claude and opencode doesn't give you this flexibility.
Learning all these stuff drove me to also set up openwebui, and it was such a successful private project that I implemented it at work (with jira/confluence/bazel query access) and management said "we need this by tomorrow".
I believe the models matter not that much anymore. The "harness" does. (unless you just want to vibe code. Thebn, throw crap at fable and call it a day)
Last night it single handedly implemented a feature after a grilling session, and came back with the red-yellow-green risk assessment points that I mostly saw with anthropic models. I had to check if I'm using the right model, but it was DS4Flash.
So maybe I was using it already?
If you are directly using it through Deepseek, they do mention that the flash slug has migrated over
> Trains: this provider may use prompts for training and may retain prompt data.
Crazy.
* DS4 Flash: 82.7
* GPT 5.6 Luna: 75.7
For reference, that puts it on the third spot behind GPT 5.5 and Fable 5. For some reason GPT 5.6 Sol is not showing in the leaderboard. If it did, then DS4 Flash was number four.
The thing is, even if Luna is better in DeepSWE and has the 80% discount. DeepSeek is still cheaper.
--------------
DeepSeek V4 | Flash GPT-5.6 Luna (New)
--------------
Input (Cache Hit) $0.0028 $0.02
Input (Cache Miss) $0.14 $0.20
Output $0.28 $1.20
--------------
Both Luna and Flash are heavy on the reasoning > output. And the cache hitrate + prices also matter.
Reality is, you can not go wrong with Luna or Flash at those prices. And remember, DeepSeek V4 Pro is still in the rafters, what is ironically closer to Luna's new price.
If those numbers translate well to its general capabilities, with the great caching DeepSeek has, I feel like this model will get tons of usage.
Although it's funny that I am thinking about that at all because I have a 2060 :P . My local inference is playing with Gemma 4 E2B and MiniCPM 5 1B.
Should be extending the lead in intelligence/cost index, as deepseek-v4-flash already were the most price efficient model, which now becomes even better. Although, in the deepseek APIs, the cost is leaking all information about codebases to China.
Does this make sense?
I didn't do much agent coding and had a mix experience.
1. It would build something that was in the spirit of what I wanted, but unusable in practice.
2. It would build something quite useful, but only the public APIs were nice, the deeper code layers would get more and more convoluted.
3. It would built what I wanted and it would have okay-ish code.
However, for 3. I also had to add a custom AGENTS.md, many more code example, extra repos as subtrees, and review any code that had new concepts.
Much more work, but still much less than typing it all by hand.
No mention of weights, just API. When will the updated weights be released?
I admire DeepSeek's openness, but even they have been raising prices after their discounts.
As for vision yeah it sucks but Luna is also 2x input and 1.5x output for 1M context...
That's around 0.4 in/1.8 out
DSv4 is wayyy cheaper.
And it's open now you have Luna at home if you have a decent set of GPUs you can run this on 2Sparks or one very expensive Mac or just like 6-8 5090s..
I guess using a ZDR provider is good enough for now.
V4 flash cache read is $0.0028 per mtok
That's not "a bit cheaper", just saying
The leaked interview has him saying it doesn't matter... as much as open source doesn't matter. There's enough in it for everyone right now and they aren't after everything.
Perspective: DeepSeek doesn't have enough infrastructure to serve their target customers already.
Rust works perfectly fine, but when I use Rust, I usually pay closer attention to performance, so I have to guide it a bit, to improve cache locality, use simd, avoid unnecessary allocations and so on. Terra has the same issues with Rust. If you always prompt models to achieve the best performance, the code is usually no the one that I want, they optimize unnecessary/cold parts or blindly optimize stuff where compiler takes care of the optimizations already.
See my other comment for more information on how I work with it. In short, keep the changes under 1k lines, context under 120k (ask it to use subagents), drive the architecture yourself.
I tried using it to generate some rust code yesterday, and it generated much code but only ever came within 1 error of a testable build. The 4th or 5th full rewrite is sitting in the buffer right now.
I'm currently looking to up my game with Bottlecap AI's return of qwen3.6, 'thinking cap'.
Alleged to be twice as fast and superior at coding over extended sessions (vs. 3.6).
We'll soon see.
I use Opus/Sol with for /brainstorming, deepseek (on Pi) for /subagent-driven-development
I love it, the docs are easily editable and when I'm ready Deepseek is faster/cheaper then everything on my coding plans.
Probably not.
But Opencode-Go is a great solution for those who don't want to pay DeepSeek directly (or can't due to reasons)
Selfish referral code: https://opencode.ai/go?ref=R1AJZT4VBX
3rd party providers on OpenRouter can be cheaper but it's already so cheap.
Otherwise, if you're trying to run it locally, even really low quantizations like DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix seem to actually not be so dumb compared to smaller models with same quantization, might be worth a try if you're sitting on a lot of RAM/VRAM yet not industry-scale amount :)
https://opencode.ai/docs/go/#usage-limits
Previously, OpenCode Go had higher API prices for some models, but now they lowered the API price and simultaneously reduced the allowance.
GPT 5.6 Luna is a new model in Go since I last checked, for example.
I still find this today:
> The plan is designed primarily for international users and provides stable global access. Your data will not be used for model training.
Played around for a few hours and used up 80 cents of tokens.
Providers' cache hit stats are available to consult, and only 1-2 of them behave properly if I remember correctly, zero if you request providers that don't store and train on sessions.
https://openrouter.ai/rankings?view=day#leaderboard-table
These days cost per task is more important, and SOTA models have become expensive.
These massive jumps in cheap models, is really great times!
I'm still considering pulling the trigger on the annual subscription of Kimi for K3 but it's sometimes slower than I'd like (at least when compared to Anthropic) even on their Vivace plan, and the token limits on the GLM Coding subscription for GLM 5.2 were too easy to hit.
Why not call it V4.1?
edit: nope, at least deepseek kept "deepseek-v4-flash" and just updated model underneath. I guess preview is no longer worth serving with that release and you'd have to look through inference provider docs to see if they've updated, yeah..
The `-0731` style suffix is worse compared to a proper version bump like V4.1.
https://news.ycombinator.com/item?id=49082022#49087112
How is that important? Maybe it does, so what?
Some providers like OpenRouter now call it `deepseek-v4-flash-0731`, but even in places like here on HackerNews people say things like "Sonnet is better than DeepSeek" without specifying a version or a reasoning effort, certainly no one will mention that `-0731` suffix when talking about DeepSeek V4 Flash.
The only issue i see, is 3th party providers that have not yet updated. But that is going to be a short time periode. There is no reason to not update.
https://github.com/antirez/ds4 is often mentioned for DS4F on Mac, but 64GB is likely not enough to achieve reasonable speeds (official weights should be ~160GB).
The max version I could order now with 128 GB?
If so, the price for local inference would be 12 000 € vs 500 000 € for a B300.
You can also do 2x 6kPRO in a workstation, for ~20k.
But still, even for mid level projects API is orders of magnitude cheaper, since you don't need to set it up and maintain it.
The 4 bit quant technically fits (there's a 127 GB version) but ... obviously that's not going to work. It is so close though, surely someone will a way to do it.
https://x.com/SourceCodeplz/status/2083099712760987746
i prefer GPT-5.6 Luna honestly
• Terminal Bench: 56.9 → 82.7 (+25.8)
• Toolathlon: 51.8 → 70.3 (+18.5)
Compared to GPT-5.6 Terra:
• Terminal Bench: Flash 82.7 vs Terra 78.4
• Toolathlon: Flash 70.3 vs Terra 53.1
• DeepSWE: Flash 54.4 vs Terra 69.6
• Agents' Last Exam: Flash 25.2 vs Terra 50.4
Trading blows with Terra, which is pretty interesting. No clear winner on these benchmarks, and wildy differeing scores. Very interesting!
Terra 87.4
https://openai.com/index/gpt-5-6/
> 78.4
The real score is always the official benchmark.
We need to see later if DS4 flash 0731 is going to maintain the score but we need to look at the official benchmarks.
Already seen a PR for DeepSWE to update the benchmark with 0731, so we can verify claimed vs official.
They're literally comparing the previous version of the same model with the new one. It's based on the same architecture, same pre-trained model, just different post-training. It doesn't get more apples to apples than this.
The performance changes are so big with the right harness that is makes sense to engineer the harness and fine-tune the model to one another from the start.
Wow that's crazy
Good times for those that don't need strict data protection