DeepSeek V4-Flash-0731: Benchmarks, Price, What Changed

The headline is not a bigger model. On Friday, July 31, 2026, DeepSeek released V4-Flash-0731, a retrained build of its lightweight open-weight model that the company says beats its own flagship preview on every agent and coding benchmark it published — while leaving the architecture, the parameter count, and the price completely unchanged. For developers already calling the API, the upgrade was silent: same endpoint, same key, same model name, better results.

The release lands one week after DeepSeek moved the broader V4 family from preview to general availability, and it reinforces a pattern that has defined the lab’s 2026: squeezing frontier-adjacent capability out of models small enough to run on a workstation, and giving the weights away under a permissive license.

What actually changed

DeepSeek describes V4-Flash-0731 as a re-post-trained version of the existing Flash model rather than a new design. The architecture is identical to the V4-Flash that shipped earlier this year: a mixture-of-experts network with 284 billion total parameters and roughly 13 billion active per token, carrying the same 1-million-token context window. Nothing about the base pretraining or the model’s size was touched.

What moved is the post-training — the reinforcement-learning and instruction-tuning stages that sit on top of the pretrained base. DeepSeek reran that pipeline with a heavier emphasis on agentic and coding tasks, the workloads where a cheap model earns its keep in production. The company published nine benchmarks and reports that the 0731 build beats the earlier V4-Pro-Preview on all nine — a notable claim, given that Pro is the larger, more expensive tier. If the technique is unfamiliar, our explainer on mixture of experts covers why sparse activation lets DeepSeek advertise a 284-billion-parameter model that still fits in a single machine’s memory.

The benchmark jumps

The gains DeepSeek reports are concentrated in exactly the categories it retrained for:

• Terminal Bench 2.1, which measures a model’s ability to operate a command-line environment, rose to 82.7 — up from 72.1 for V4-Pro-Preview and 61.8 for the earlier Flash Preview.
• DeepSWE, an agentic software-engineering test, climbed to 54.4, a step-change from the 7.3 the preview build posted on the same benchmark.
• Cybergym, a security-focused agent evaluation, came in at 76.7.

On the independent side, Artificial Analysis scored V4-Flash-0731 at 50 on its Intelligence Index, placing the model among the top three open-weight systems on that leaderboard and creating what the firm described as a sharp spike on the price-performance Pareto frontier. Vendor-published benchmark deltas always warrant caution — the meaningful test is whether the numbers survive independent reproduction, which is why we keep returning to the discipline of a proper LLM evaluation. But an outside index landing the model in the open-weight top tier is a harder data point to wave away.

Same price, same endpoint

The economics are the point of a Flash-tier model, and DeepSeek held the line. V4-Flash-0731 ships at $0.14 per million input tokens (cache-miss) and $0.28 per million output tokens — identical to the pricing the earlier Flash build carried. There is no premium for the retrained version.

The migration story is unusually clean. Developers calling the deepseek-v4-flash model ID received the new build automatically: same endpoint, same API key, same model name, zero code changes. DeepSeek also confirmed the API now natively supports the Responses API format and is adapted for coding-agent frameworks, lowering the friction for teams wiring the model into an AI agent or an existing tool-calling harness.

Open weights, MIT license

Consistent with the rest of the V4 line, DeepSeek published the weights on Hugging Face under the MIT license, which permits unrestricted commercial use and modification. That matters for the audience most likely to care about a cheap, capable Flash model: teams that want to self-host rather than route sensitive traffic through a Chinese-operated API.

At 284 billion total parameters with only 13 billion active, the model is small enough — after quantization — to run on high-memory consumer hardware, and early community guides have documented it fitting on a single 128 GB machine. Our primer on quantization explains the tradeoffs that make that possible: shrinking the weights’ numerical precision to cut the memory footprint, ideally without gutting quality.

The open-weight cadence continues

V4-Flash-0731 arrives in the densest stretch of open-weight releases the industry has seen. It follows DeepSeek’s own V4 general-availability launch a week earlier and Moonshot AI’s Kimi K3, and it slots into a now-familiar dynamic: Chinese labs operating under U.S. export controls keep shipping frontier-adjacent systems and releasing the weights, tightening the squeeze on Western providers at the low end of the price curve.

That pressure has been showing up in pricing as much as in benchmarks. Days before the 0731 build, OpenAI cut the price of its GPT-5.6 tiers sharply in what our coverage framed as an escalating AI price war. A retrained open-weight model that closes part of the capability gap without raising its price is another turn of the same screw, and it reinforces the through-line of our reporting on open-source models closing the gap: the distance between the best closed and best open systems keeps shrinking, and it is shrinking fastest at the value tier.

What it means

The interesting part of this release is what it says about where the returns now come from. DeepSeek did not spend more compute on a bigger model; it spent it on better post-training of a model it already had — and reports flagship-beating agent scores as a result. For a field that has spent two years equating progress with scale, a same-size, same-price model that jumps double digits on Terminal Bench is a reminder that the post-training stack still has a lot of slack in it. Expect every lab that ships a Flash-class tier to be studying how much of its own headroom is sitting unclaimed in RL and tool-use tuning.

Who benefits. Cost-sensitive teams running high-volume agentic or coding workloads get a free capability bump — the silent upgrade means anyone already on deepseek-v4-flash woke up on August 1 with a stronger model at the same bill. Teams that self-host the MIT-licensed weights get the same gains without depending on DeepSeek’s API at all, which is the deployment path most enterprises with data-residency concerns will actually take.

The caveats that remain. The nine benchmarks are DeepSeek’s own, and the categories that improved are precisely the ones it retrained for — the risk with targeted post-training is that gains on the measured tasks do not fully generalize to messier real-world work. Artificial Analysis corroborating the top-tier placement helps, but the honest read is to treat the flagship-beating claims as promising until independent agent evaluations confirm them. And for regulated buyers, a Chinese-origin model — however permissively licensed — still carries provenance questions that a benchmark cannot answer.

What to watch. Whether independent evaluations reproduce the Terminal Bench and DeepSWE jumps; how quickly the retrained weights show up in self-hosted deployments now that the capability-per-dollar has moved; and whether Western labs answer a same-price capability gain with more price cuts, more capability, or both. The competitive frontier for cheap, capable models is now being contested on post-training quality as much as on raw scale — and DeepSeek just moved it again without shipping a single new parameter.