跪拜 Guibai
← Back to the summary

DeepSeek V4 Flash Ships: 13B Active Parameters, 6x Agent Jump, and a Price Floor of $0.003/M Tokens

On the afternoon of July 31, DeepSeek held no press conference, posted no teaser tweets, and simply updated a line in the API documentation: the official version of DeepSeek-V4-Flash (V4-Flash-0731) is now in public beta.

But this understated step has completely broken new ground.

A "budget model" with only 13 billion activated parameters and a price as low as 0.02 yuan per million tokens is crushing the V4-Pro preview version from three months ago in multiple Agent benchmarks. image.png

First, the most explosive number: DeepSWE surged 6x

Compared to the Preview version, the official V4-Flash has exactly the same model structure and parameter scale—total parameters 284 billion (MoE architecture), activated parameters 13 billion. The underlying base model was untouched; only a new round of post-training was conducted.

But Agent capabilities achieved a qualitative leap. The officially announced benchmark scores are as follows:

Benchmark V4-Flash Official Score
Terminal Bench 2.1 82.7
NL2Repo 54.2
Cybergym 76.7
DeepSWE 54.4
Toolathlon verified 70.3
Agent Last Exam 25.2
Automation Bench (Public) 25.1
DSBench-FullStack 68.7
DSBench-Hard 59.6

The most shocking is DeepSWE—a benchmark specifically designed to evaluate an AI Agent's ability to autonomously solve problems in real, long-cycle, multi-step programming tasks. The Flash official version's score jumped from 7.3 in the preview version to 54.4, a more than 6x increase.

Look at Terminal Bench 2.1 (an authoritative benchmark evaluating an Agent's ability to autonomously plan and execute multi-step command-line tasks and error recovery in a Linux terminal environment): the Flash official version scored 82.7, higher than V4-Pro-Preview's 72.1 and GLM-5.2's 81.0, approaching Opus-4.8's 85.0.

In the intelligence index test by the overseas evaluation agency Artificial Analysis, V4-Flash-0731 (highest reasoning tier) scored 50 points—10 points higher than the V4-Flash released in April, 6 points higher than V4-Pro, and only 1 point lower than GPT-5.6 Luna's highest score of 51.

The most chilling point: The structure didn't change, it was just "re-educated"

The most noteworthy aspect of this upgrade is not the scores, but how the scores were achieved.

V4-Flash-0731 has exactly the same model structure and parameter scale as the preview version. The underlying base model was not altered; only a new round of post-training optimization was completed. In the words of NetEase Youdao: the base hasn't changed; what changed is the "acquired education."

This means that over the past three months, the DeepSeek team, through more refined post-training strategies (a two-stage paradigm of expert model cultivation + policy distillation), significantly improved the model's tool-calling ability and multi-step task execution capability without changing the architecture. For the same task complexity, the new model can complete the same work with fewer token consumptions.

What does this imply? The room for optimization in post-training might be much larger than we imagined. This also explains why DeepSeek dares to say "the official version of V4-Pro will be released soon"—applying the same post-training method to Pro will only yield more exaggerated results.

Technical Foundation: Million-token context, efficiency revolution

The core parameters of V4-Flash are already very impressive:

Underpinning all this are three technological innovations of the V4 series:

  1. Hybrid Attention Architecture (CSA + HCA): Traditional Attention has O(n²) complexity in long contexts. V4 introduces a compression mechanism at the token dimension—reading closely related tokens carefully while compressing or skipping weakly related tokens, breaking through the quadratic complexity bottleneck.
  2. Manifold Constrained Hyperconnection (mHC): Solves numerical instability during deep stacking, improving training stability by 6.7%.
  3. Muon Optimizer: Full-chain inference acceleration up to nearly 2x.

In a 1 million token scenario, V4-Flash's single-token inference FLOPs are as low as 10% of V3.2's, and KV cache usage is as low as 7%. The pain points of long-text scenarios—"can't run, can't remember, can't afford"—have been directly shattered.

https://api-docs.deepseek.com/quick_start/agent_integrations/codex/

Price butcher: 0.02 yuan/million tokens, 98% cache hit discount

V4-Flash's pricing remains a "butcher's price":

Billing Item Regular Price Peak Hours (9:00-12:00, 14:00-18:00)
Input (Cache Hit) 0.02 yuan/M tokens 0.04 yuan/M tokens
Input (Cache Miss) 1 yuan/M tokens 2 yuan/M tokens
Output 2 yuan/M tokens 4 yuan/M tokens

To put it in perspective: processing a 100,000-character article with Flash costs less than 1 cent. Some media outlets calculate its price is only about 1/90th of Claude's.

More critical is the caching strategy—DeepSeek offers approximately a 98% cache hit discount for its own API, far exceeding the 90% offered by most vendors in the industry. The input price differs by 50 times between a cache hit and a miss. Some developers have tested and found that after restructuring their caching strategy, the cost per request dropped from 1.2 yuan to 0.17 yuan, a direct 85% cut.

For high-frequency Agent workflows, RAG applications, and batch generation scenarios, this price means the barrier to enterprise-grade AI capabilities can be pushed down to almost negligible levels.

The ecosystem is already moving

Just after the official version launched, the ecosystem reacted extremely quickly:

Who should use Flash?

A one-sentence summary of the V4 dual-version positioning: Choose Pro for heavy reasoning, long documents, and complex automation tasks; choose Flash for high concurrency, low cost, and batch light tasks. On simple Agent tasks, the Flash official version is on par with Pro.

Specifically, Flash is suitable for:

And the official version of V4-Pro is still on the way—DeepSeek officially stated that this upgrade only applies to the V4-Flash API; the V4-Pro API and APP/WEB end models have not been changed.

Conclusion

Can't get top-tier chips? Then rely on engineering optimization to blow the price through the floor—this is very DeepSeek.

From the open-sourcing of the V4 preview version in April to the official Flash version at the end of July, this company has been impressing people with the same story: not stacking parameters, but squeezing every bit of computing power and every penny to the extreme. 1M context, 13B activated parameters, a 0.02 yuan price, near-top-tier Agent capabilities—when these four terms appear on the same model, the industry's pricing system is about to be reshuffled again.

The day the official version of V4-Pro is released will likely be the real "trump card" moment.