On July 31st, DeepSeek made a move that caught the entire industry's attention: they open-sourced the official version of V4 Flash (under the MIT license). This model, activating only 13B parameters(the part of the model actually used each time), scored 50 points on a third-party intelligence index—only 1 point shy of OpenAI's closed-source flagship, GPT-5.6 Luna.

Think of this as a delivery station: inside, there's a small electric scooter (Flash, 13B activated) and a big truck (Pro/closed-source flagship). The truck has a larger engine and can go further, making it the ace. One day, the scooter got a new navigation algorithm—without changing its engine or battery, just updating the software—and it started delivering meals faster than the truck. The kicker? This navigation algorithm is now freely available for anyone to install. The analogy ends here, but the real difference is: the truck driver won't release the engine blueprints, but DeepSeek really did release all the weights(the model's core files — get them and run it yourself).
Event

No Skeleton Change, Just Training Update, Flash Outperforms Its Own Pro

DeepSeek officially stated: the model structure and size of V4 Flash 0731 are exactly the same as the preview version from three months ago—total parameters 284B, activated 13B, 1M context window(how much text it can hold at once). The only difference is a new round of post-training(keeping the model skeleton unchanged, only adjusting the "how to answer questions" strategy, similar to giving the same person a new work method).

So what happened? According to the official announcement of 9 Agent benchmark tests, the official version of Flash completely surpassed the higher-positioned V4-Pro preview version. The most astonishing was DeepSWE (high-difficulty coding): jumping from the preview version's 7.3 to 54.4, an increase of more than 6 times. Terminal Bench 2.1 (terminal operation) scored 82.7, surpassing Pro's 72.1, and even outperforming ZhiPu GLM-5.2's 81.0, just 2.3 points shy of Claude Opus 4.8's 85.0.

In other words: the industry's default "Pro is always stronger than Flash" hierarchical logic has been broken by DeepSeek itself. The lightweight version is not a low-configuration version—it just hasn't been trained well enough.

82.7
Terminal Bench 2.1
Terminal operation test, surpassing Pro preview by 10.6 points, just 2.3 points shy of Opus 4.8. Source: DeepSeek official.
54.4
DeepSWE (preview version 7.3)
High-difficulty coding, increased by more than 6 times compared to the preview version. Source: DeepSeek official.
70.3
Toolathlon Verified
Comprehensive tool invocation test, surpassing Pro preview by 14.4 points. Source: DeepSeek official.
Numbers

1 Point Away from Closed-Source Ceiling, 60% Cheaper

The third-party evaluation agency Artificial Analysis's intelligence index(a comprehensive ability score across multiple tasks, with a comparable model median of about 17 points) gave V4 Flash 0731 a 50 points. What does this number mean?

Horizontally: on par with Google Gemini 3.6 Flash (50 points), 1 point shy of GPT-5.6 Luna (51 points), 7 points away from the current open-source leader Kimi K3 (57 points). Vertically: compared to its own Flash preview version, it improved by 10 points (40→50), and compared to the higher-positioned V4-Pro preview version, it is 6 points higher (44→50). This is the first time in open-source model(models whose files are public for anyone to download and build on) history to come so close to this level of closed-source flagship.

Explosive Cost Performance

Price-wise: cache hit input 0.2 yuan/million tokens(the units AI bill by), cache miss input 1 yuan, output 2 yuan. Calculated by task cost, it is about 60% cheaper than the reduced-price GPT-5.6 Luna. OpenAI reduced Luna's price by 80% on the same day, but it is still more expensive than DeepSeek.

ModelAA Intelligence IndexActivated ParametersOpen SourcePer Task Cost
V4 Flash 07315013BMITBaseline
GPT-5.6 Luna (max)51Not disclosedNo+60%
Gemini 3.6 Flash50Not disclosedNo
Kimi K3 (max)57Not disclosedYes
V4 Pro preview version4449BNoHigher

Data source: Artificial Analysis Intelligence Index (2026-07-31), DeepSeek official pricing page. Per task cost is calculated according to the Artificial Analysis standard.

Counterintuitive

Post-Training More Effective Than Parameter Stacking—Who's It Most Beneficial For?

The most counterintuitive aspect of this release: the model skeleton remains completely unchanged. 284B total parameters, 13B activated, CSA+HCA hybrid attention, 32T token training data—all the same as three months ago. The only change is the post-training strategy.

What does this mean? The current performance bottleneck of large models is likely not in "how big the model is," but in "how refined the training method is." The same skeleton, with a different post-training strategy, can achieve a qualitative leap—this is a huge boon for teams with limited computing power(the computing resources needed to run models): you don't need more GPUs, you need better training methods.

DeepSeek's official statement also hinted at a signal: the official version of V4-Pro will be launched in early August. If Pro also undergoes the same post-training upgrade, the ceiling of the open-source flagship will be raised even further.

Why "Just Training Change" Can Lead to Qualitative Change
01
Skeleton Unchanged
284B/13B MoE, 1M context, architecture consistent with the preview version
02
Post-Training Redone
More refined RL/alignment strategy, strengthened for Agent scenarios
03
Agent Ability Qualitative Change
9 benchmarks all surpass Pro, DeepSWE 6 times increase
04
Open Source Release
MIT license, 167GB weights, anyone can download and deploy
Core Insight: Performance bottleneck is not in parameter quantity(the number that measures a model's size), but in training method. Limited computing power ≠ intelligence ceiling.
Landscape

The Balance Between Open Source and Closed Source Is Shifting

Taking a broader view. What happened in the first half of 2026? Kimi K3 was so popular it was sold out, DeepSeek V4 Flash caught up with closed source, GLM-5.2 followed closely behind—Chinese open-source models are collectively approaching the global ceiling.

Meanwhile, what is the closed-source camp doing? OpenAI reduced the price of GPT-5.6 Luna by 80% on the same day. This is not a coincidence—when open-source models are only 1 point behind on the intelligence index and 60% cheaper, the only moat left for closed-source vendors is the convenience of "you don't have to deploy it yourself."

But the MIT license means: any team can download the weights, privatize the deployment, and fine-tune(retraining it on your own materials so it fits you) according to their own needs. For data-sensitive enterprises (finance, medical, government), this is not a question of "whether to save money," but "whether it can be used." Open-source models have for the first time simultaneously met the two conditions of "smart enough" and "able to be privatized."

MIT
License
Unrestricted commercial use, modification, and redistribution. Highly attractive to startups and data-sensitive enterprises.
167GB
Weight File Size
Released in FP4/FP8 mixed precision, including DSpark draft module. Q8 quantization(slimming the model file so it runs on less hardware) can run with about 110GB memory.
Early August
V4-Pro Official Version
DeepSeek official revealed that the Pro official version will be launched soon, and will also integrate Codex.
Action

Available Today: Who Should Switch and How

V4 Flash 0731 has been launched on DeepSeek's first-party API, and Ollama has also listed it simultaneously (ollama run deepseek-v4-flash:0731-cloud). It natively supports the OpenAI Completion API format and is adapted for Codex.

Should You Switch? Three Judgments
1

If you have a large number of repetitive Agent tasks (code generation, terminal operation, tool invocation): prioritize testing. Flash 0731 has already approached Opus 4.8 in these tasks, with a cost 60% lower.

2

If you need private deployment (data does not leave the domain): you can download it now. MIT license + 167GB weights + Q8 quantization can run with 110GB memory.

3

If you need top-level planning/creative ability: wait a bit longer. V4-Pro official version will be launched in early August, with a higher positioning. Flash excels in "doing work," not "giving ideas."

Quick Start
APIapi.deepseek.com, model name deepseek-v4-flash, compatible with OpenAI SDK
Localollama run deepseek-v4-flash:0731-cloud (or download GGUF to run offline)
CodexNatively adapted, configure DeepSeek as the provider in Codex CLI / VS Code plugin
PriceInput 1 yuan/million tokens (cache hit 0.2 yuan), output 2 yuan/million tokens

When a 13B activated open-source model is only 1 point away from the closed-source ceiling and 60% cheaper, the strategy of "waiting for closed-source price reduction" is outdated. The first thing you can do today: run an A/B test on the most repetitive part of your Agent workflow with V4 Flash. The data will make the decision for you.

Benchmark data comes from DeepSeek's official release and third-party evaluations by Artificial Analysis. Some benchmarks (DSBench series) use DeepSeek's self-developed Harness, which is not the same set of questions as the old version of Terminal Bench 2.0, and the actual production environment performance still needs third-party reproduction.