The article states you can fit Q4 in 4 x 4090 and it works reasonably well.
I'd personally fo for deepseek V4 flash at Q8, hardware prices need to come down though. Once an NV4 version get released it'll be easier to run on commodity hardware.
Open-weight models with strong multilingual support change the math because you can self-host at marginal cost once you have GPU capacity. DeepSeek's earlier versions already punched above their weight on non-English benchmarks (especially CJK and some Indic languages where the gap to GPT-4 was much narrower than English-only benchmarks suggested).
Two questions for anyone who's actually deployed V4 in production yet:
1. How does it handle Turkish / Slavic morphology compared to V3? In our tests V3 was solid for Russian and respectable for Turkish, but handled compound morphology in agglutinative languages a bit awkwardly.
2. Is the long-context window actually usable end-to-end or does quality degrade past ~64k like with most open models?
Having said that I really hope this model of deepseek, performs significantly on par with the claude saunnet model.
I believe Claude Code only works with claude and seems all I hear about that is it's great but the token limits are so anemic as to make it useless unless you want to shell out $200+ a month, which I do not so I haven't bothered.
I tried codex but it wouldn't run out-of-the-box. Installation on a fresh Windows box resulted in some obscure error which is a strong "this product isn't fully baked" signal.
Open code desktop thus far has been the only turn-key solution, worked right away on it's pickle model but was a real pain to hook to anything else. It exhibits a lot of the typical obtuse UX that open source projects end up with since open source tends to attract coders-developers more than UX/UI people. At least it does mention that it's still beta.
I would say DeepSeek is definitely behind compared to Codex but Claude doesn’t and hasn’t impressed me for some time now. It writes way too much code when it doesn’t need to in a fashion that gradually rots your codebase.
Codex is the only model I’ve used which will regularly remove more code than it adds or make a fix or feature by adding a single line of code or otherwise do minimal working changes.
Claude is the model which can get the feature working by adding two new classes, 20 new methods and 2000 lines of code, when it actually needed to remove 500 lines of code and add two new methods.
Claude will also often refactor by adding tons of new code and using it while not deleting any of the old code.
But that was 3 months ago, have not tried it since, they could have grown.
To be fair, I think what you are meaning, if I drop the literal frame here, is this, tell me if I am right:
Codex > Claude in my setup.
that right?
To be fair, my tests were not apples to apples. I have sophisticated agent alignment harnesses which prevent claude from hallucinating or going off the rails, ( not literally, not 100% = about 80% less hallucination, about 90% less drift, and about 98% more starting from crystal clear intent.
And in my personal tests, codex was not calibrated to use those systems, it had them but would have needed to find them.
Also I am in a massive project, next ai labs, ixcoach, with likely in the range of 20k files of code, 100x files of docs...
It could just be my agent alignment harness thats making claude outperform codex. Looking into testing it on the major benchmarks and publishing the results.
There is a fun term “jagged frontier”.
Meaning: one model can be much better than the other one in one thing, and much worse than the other in another thing.
Codex 5.5 extra high currently feels a good amount smarter than either 4.6 or 4.7 Opus. I only just started using it about a week ago, so maybe that's a recent development and then OpenAI will eventually lobotomize their model or throttle etc.
What I dislike about frontier models is how opaque and incentivized the businesses are about tweaking their services. Anthropic definitely does some shady throttling. I have zero trust for Altman and he's BS AGI claims. And Google makes it non-obvious that you can't turn Gemini training off on even their highest tier personal plan. There's a lot of shady and dishonest behaviour, probably because they are all overhyped and heavily subsidized to win the race. I don't mind at all paying more than I currently am for these services, but I don't trust any of these frontier model companies, and so I'm cheering for open models.
Right now I'm using Codex for planning and DeepSeek V4 Flash [1m] for implementation. It's quite fast. Quite possible / likely that OpenAI will make significant changes that kill this workflow for w/e reason... at which point I will probably move to full open weight models.
I use Claude Code with GLM 5.q, Kimi K2.6, MiniMax M2.7 and Xiaomi MiMo V2.5 Pro.