DeepSeek releases V4-Flash-0731 for agents, but not its evaluation harness

On July 31, DeepSeek released DeepSeek-V4-Flash-0731, the official version of its Flash model for agents. It is not a new architecture: it retains the April preview's design and, according to the company, has only undergone new post-training. It is nevertheless a material change in capability and availability. A new MIT-licensed checkpoint is available, the public API now serves it under the name deepseek-v4-flash, and the model supports the Responses format used by several coding agents.

The distinction from April's launch matters. The V4 technical report describes a mixture-of-experts model with 284 billion total parameters, 13 billion active per token, and a one-million-token context window. The 0731 version retains that base, adds the DSpark speculative-decoding module to the checkpoint, and puts the novelty in how the model was trained after pretraining to reason and act with tools.

In other words, DeepSeek did not need to present another model scale to claim another level of behavior. That is the most interesting part of the release.

The improvement appears in work loops, not isolated questions

The V4-Flash-0731 model card compares the new checkpoint with the preview on nine agent evaluations. On Terminal-Bench 2.1, DeepSeek reports a rise from 61.8% to 82.7%; on Toolathlon-Verified, from 49.7% to 70.3%. It also reports large improvements in software engineering, cybersecurity, and multi-step professional tasks.

Those two tests help explain the claim without turning this entry into a table. Terminal-Bench 2.1 contains 89 verifiable tasks inside a terminal: configuring software, processing data, compiling projects, or solving scientific problems. Its own May revision fixed 28 tasks because external dependencies, resource limits, or faulty instructions were changing results. Toolathlon-Verified presents 108 tasks combining 604 tools across 32 applications, with roughly twenty interaction rounds on average and executable graders.

A gain on both is therefore more informative about persistence and tool use than another improvement on a short-answer exam. It does not, however, show that the model can complete 82.7% of “terminal tasks” in general, or that it will retain the score on a particular organization's repositories, tools, and policies.

There is a more immediate limitation: for coding tests, DeepSeek used the max reasoning level and the minimal mode of DeepSeek Harness, which the company says will be released later. Two of the nine tests are also internal sets. The company has published the weights that generate tokens, but not the entire system that chose what context to provide, when to run a tool, how to observe its result, and when to stop.

That currently prevents anyone from separating how much of the gain comes from post-training and how much comes from the scaffold. The figures also lack task-level cost, latency, retry counts, and full trajectories. The benchmarks have verifiable tests, but this particular result remains a provider measurement under a configuration that another party cannot yet replay end to end.

Open does not mean small or complete

The checkpoint does enable much deeper scrutiny than a closed API. Its weight index lists 48 shards occupying roughly 167 GB; the repository also includes the message encoder and speculative module, and is distributed under MIT. A lab can inspect, modify, and run it without depending on DeepSeek to keep an endpoint alive. But this is not a laptop model: the model card's featured recipe uses a node with four GB300 GPUs, and the vLLM recipes show how memory, context length, and parallelism constrain which modes are practical.

The API removes that operational barrier. Its pricing documentation currently lists $0.14 per million uncached input tokens and $0.28 per million output tokens, with a maximum output of 384,000 tokens. Those are very low rates for experimenting with long-running agents. They are not the cost of a solved task: max mode may deliberate at length, repeat actions, or fail at the end, and DeepSeek does not pair its scores with token use and time. The page also says rates will double during certain peak hours on a date yet to be announced.

For software builders, the practical change is real. They can replace the backend of a Responses-compatible agent, test the checkpoint on their own infrastructure, or design a harness that constrains tools, permissions, and budgets. But the release also makes clear that “using V4-Flash-0731” does not by itself identify a reproducible system. At minimum, the weight version, reasoning level, message format, tools, execution loop, context limits, and evaluator must be pinned.

My reading is that this is the release's durable lesson: post-training and the harness are becoming part of model capability. If the same architecture jumps twenty points across different tests, the useful question is no longer only how many parameters it has, but what action experience it received and what system accompanies it during measurement. Publishing the checkpoint is necessary for honest agent comparisons, but so are the scaffold, versions, trajectories, and cost.

Openness does not settle safety either. The 0731 card includes a cyber-capability score but no checkpoint-specific evaluation of misuse, refusal, action control, or incidents. With MIT-licensed weights, any safeguards in DeepSeek's service can be replaced. That does not turn the model into an attack, but it does move responsibility for tool isolation, least privilege, and behavioral testing to the operator—areas a completion table does not measure.

V4-Flash-0731 deserves attention even if its scores are less portable than the announcement suggests. The weights exist, the API is live, and the post-training gain claimed on tasks with executable outcomes is large. The next important test will not be another number from DeepSeek. It will be the lab releasing its harness so others can determine how much of that capability actually travels with the model.

Sources

• DeepSeek, DeepSeek-V4-Flash change log, July 31, 2026.
• DeepSeek, DeepSeek-V4-Flash-0731 checkpoint and model card, accessed August 4, 2026.
• DeepSeek-AI, DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence, April 26, 2026.
• DeepSeek, API models and pricing, accessed August 4, 2026.
• Terminal-Bench, Terminal-Bench 2.1, May 6, 2026.
• HKUST NLP, Toolathlon-Verified, June 30, 2026.
• vLLM, deployment recipes for DeepSeek-V4-Flash, accessed August 4, 2026.