DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains
MarkTechPost Asif Razzaq ● Covered by 6 sources
DeepSeek just retrained its V4-Flash model, not rebuilt it, and it now beats the pricier V4-Pro on agentic benchmarks. Same size, smarter training, and it costs a third as much to run via API.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DeepSeek dropped V4-Flash-0731 on Hugging Face this week and quietly flipped the switch on its public API beta, and the interesting part isn't the model itself so much as what it implies about where the easy gains in AI are hiding right now. The architecture hasn't budged an inch since the April preview: still a 284-billion-parameter mixture-of-experts model with 13 billion active per token, still a 1M-token context window. DeepSeek didn't touch the design. They just re-ran post-training, and the result apparently now beats their own V4-Pro across every agentic benchmark they published, from Terminal Bench 2.1 to DeepSWE.
That's a notable admission, whether DeepSeek intended it that way or not. Post-training tweaks closing the gap with a supposedly more capable sibling model suggests the frontier of usable performance right now is less about parameter count and more about how you teach a model to use tools, follow agent loops, and not fall apart on messy multi-step tasks. V4-Flash-0731 posts a 54.4 on DeepSWE versus 12.8 for V4-Pro's preview, and a jump from 38.5 to 54.2 on NL2Repo. Those are not rounding errors.
Pricing is where this gets practical fast. DeepSeek is charging $0.14 per million input tokens on a cache miss and just $0.28 per million output tokens through the API, which is roughly a third of what V4-Pro output costs. For a solo developer or a startup running agent loops all day, that's the difference between a side project and a real product. No GPU cluster required, no procurement meeting, just an API key.
Self-hosting is a different story entirely. The weights are MIT-licensed and ungated, which is genuinely generous, but DeepSeek's own MoE design means every expert sits resident in memory even though only 13 billion parameters fire per token. Unsloth's quantized builds still need around 110GB of combined RAM and VRAM at 3-bit, and DeepSeek's reference setup for full precision is a 4x GB300 node — hardware that lives in enterprise racks, not laptops. So the MIT license mostly benefits companies that already have serving infrastructure, not hobbyists tinkering on a single GPU.
One caveat worth sitting with: every benchmark number here comes from DeepSeek, run on an internal harness called DeepSeek Harness that hasn't been released publicly. Agent benchmarks are notoriously sensitive to harness details — prompt formatting, retry logic, tool-calling conventions — so an outside lab replicating these numbers with a different scaffold could land somewhere quite different. The gains look real, but they arrive without a way for anyone else to independently check the math yet.
My take — AI-written commentary, not fact-checked reporting
I like that DeepSeek is proving you can squeeze real agentic gains out of post-training alone, and I like even more that they're shipping it MIT-licensed instead of hiding behind an API wall like most Western labs would. But let's not pretend vendor-reported numbers on an unreleased harness are the same as verified results — until someone outside DeepSeek reruns these evals, treat the leaderboard bragging rights as provisional, not gospel.
Read more about this at: MarkTechPost
