DeepSeek has updated again, but this time Liang Wenfeng didn't drop a big move.

Liang Wenfeng's popularity has remained top-tier in the AI circle, especially this year.

From the release of the V4 preview version, to subsequent financing and IPO rumors, to a widely circulated four-hour conversation transcript, every move by DeepSeek quickly becomes an industry topic.

DeepSeek updated again.

This time, DeepSeek only announced in the update log of its API documentation that the V4-Flash official version API has launched public beta.

From the simultaneous appearance of the words “official version” and “public beta,” this is not V4's final form. DeepSeek also specifically emphasized that this update only involves the V4-Flash API; the V4-Pro and the models on the App and Web ends have not changed, and the V4-Pro official version will need to wait a while longer.

But this update is still worth paying attention to.

Looking at the latest version, V4-Flash-0731 did not change the model architecture or parameter scale, only undergoing another round of post-training. After the update, DeepSeek devoted almost all of its text to emphasizing its progress in Agent, especially Code Agent.

This means that V4-Flash is not a simple model update.

DeepSeek is trying to answer a new question: at the Agent stage, does a company most need a model with the strongest capabilities, or a model that is strong enough, fast enough, and cheap enough?

This time, DeepSeek did not continue stacking parameters

This is not V4-Flash's first appearance.

When DeepSeek released the V4 preview version, it had already launched V4-Pro and V4-Flash simultaneously. Both are MoE models, but the size gap is significant: V4-Pro has 1.6 trillion total parameters, activating 49B parameters per inference; V4-Flash has 284B total parameters, activating only 13B parameters per inference. Both models support 1 million Token context.

In DeepSeek's positioning, Pro is responsible for capability ceilings, and Flash is responsible for efficiency.

The larger V4-Pro performs better on world knowledge and high-difficulty Agent tasks; V4-Flash can approach Pro on some reasoning tasks by increasing the thinking budget. In other words, Flash sacrifices some knowledge capacity but uses smaller activation parameters in exchange for speed and cost.

This update did not change this architecture.

DeepSeek clearly stated that V4-Flash-0731 has exactly the same model structure and scale as the preview version, with only another round of post-training performed.

But it was precisely this post-training that brought obvious changes to V4-Flash's Agent capabilities.

According to the results published by DeepSeek, V4-Flash-0731 reached 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 54.4 on DeepSWE, and 70.3 on Toolathlon Verified. DeepSeek said these results significantly surpass the V4-Pro preview version.

The significance of this is that DeepSeek did not improve the model's Agent capabilities by adding parameters and re-pretraining.

In plain terms, although DeepSeek continues to pursue AGI, it is not achieving this by blindly increasing parameters, but rather by letting the model learn to better break down tasks, call tools, execute code, and reduce errors over longer task trajectories.

In the past, the main competition among large model companies occurred in the pre-training stage. Whoever had more computing power and more data had the opportunity to train models with larger parameters and more knowledge. But after entering the Agent stage, model capabilities are no longer entirely determined by pre-training.

Whether a model can use a terminal, maintain task state across hundreds of interactions, correct itself after errors, and know when to call which tools — these capabilities increasingly depend on post-training, reinforcement learning, and execution frameworks external to the model.

The change in V4-Flash precisely illustrates this point.

This is also why DeepSeek, in addition to updating the model this time, added native support for the Responses API and specially adapted Codex. Meanwhile, when testing Code Agent tasks, DeepSeek used the not-yet-officially-released DeepSeek Harness minimal mode.

In other words, these Agent results do not belong entirely to the model itself, but to a system composed of “model, reasoning budget, and Harness.”

This of course also means that the currently published data should still be viewed with caution.

On one hand, some results come from DeepSeek's internal test sets; on the other hand, DeepSeek Harness has not been made public, so the outside world cannot reproduce these results in exactly the same environment. How much the official V4-Flash has actually improved over the preview version still requires independent evaluation and validation in real development scenarios.

But at least from this update, it is clear that DeepSeek is taking a step beyond merely providing models and APIs toward the execution environment that Agents need.

But DeepSeek has not yet built a complete Agent product itself; instead, it has begun providing the interfaces, adaptations, and Harness necessary for models to enter Agent workflows.

This is a very restrained form of productization.

Why Flash?

Since V4-Pro has a higher capability ceiling, why doesn't DeepSeek update Pro first?

The most direct reason is that an Agent is not a single model call.

In the chatbot stage, a user asks a question, the model generates one answer, and the interaction ends. But in Agent tasks, the model needs to first understand the goal, then break down the task, retrieve information, call tools, read results, correct errors, and finally complete the delivery.

A complex task may involve dozens or even hundreds of model calls.

At this point, the increased cost and latency of a single model inference are continuously amplified across the entire task trajectory. A model with stronger capabilities but slower speed and higher cost may not necessarily bring better actual results.

What Agents truly need is often not calling the strongest model at every step, but using a model with sufficient capability, lower price, and faster response in the vast majority of steps.

This is precisely V4-Flash's value.

According to DeepSeek's technical report, under 1 million Token context, the computation required for V4-Flash's single-Token inference is about 10% of V3.2's, and the KV Cache is about 7% of V3.2's. Its total parameters are only about one-sixth of V4-Pro's, and its per-inference activated parameters are only about one-quarter of Pro's.

This gap is ultimately also reflected in API pricing.

Currently, V4-Flash's cache-miss input price per million Tokens is 1 yuan, and the output price is 2 yuan; V4-Pro's are 3 yuan and 6 yuan respectively. V4-Flash's concurrency limit is 2500, while Pro's is 500. The Responses API is currently only supported by V4-Flash; V4-Pro is expected to be integrated by early August.

In V4's product system, Pro is responsible for proving where DeepSeek's model capabilities can reach, while Flash is responsible for undertaking truly high-frequency, long-chain Agent tasks. The former is a capability model; the latter is more like a production model.

Liang Wenfeng once said openly in an internal setting, “Models should be compared under the same cost; at this stage, the priority of Coding Agent is higher than other vertical Agents.”

From this perspective, the direction of this V4-Flash update indeed echoes these two judgments.

DeepSeek did not choose to expand parameters to make Flash catch up with Pro in all capabilities; it chose to improve Flash's ability to complete real tasks through post-training and Agent frameworks while keeping the model architecture unchanged.

Behind this is a more realistic calculation.

If Agents are ultimately to enter the daily workflows of enterprises and developers, then what determines whether a model can be used at scale is not just Benchmark scores, but how much time it takes to complete a task, how many Tokens it consumes, and what the final success rate is.

However, this is also what V4-Flash most needs to prove at present.

Most of the Agent results published by DeepSeek were completed using Max reasoning effort. Higher thinking intensity usually means longer reasoning processes and more Token consumption. A model's low price per million Tokens does not mean the total cost of completing a task will necessarily be lower.

If V4-Flash needs to consume more Tokens to achieve results close to Pro, its advantage in unit price may be partially offset by longer task trajectories.

Therefore, a truly valuable comparison is not whether V4-Flash surpasses someone on a particular Benchmark, but how many Tokens, how much time, and how many retries it needs compared to V4-Pro, Kimi, GLM, and closed-source models when completing the same Agent task.

DeepSeek has not yet provided this answer.