跪拜 Guibai
← Back to the summary

DeepSeek V4 Flash Beats Its Own Pro Model on Agentic Tasks — and It's All Post-Training

Why DeepSeek V4 Flash Can Surpass V4 Pro Preview

On July 31, 2026, DeepSeek released the official V4-Flash-0731 version. Shockingly, this lightweight model with only 284B total parameters (13B activated) comprehensively surpassed its own V4-Pro-Preview with 1.6T total parameters (49B activated) on Agentic benchmarks. This article deeply analyzes the technical reasons behind this "inferior defeating superior" phenomenon.

Full Summary

DeepSeek V4-Flash-0731 comprehensively surpasses V4-Pro-Preview on 9 Agentic benchmarks including Terminal Bench 2.1, DeepSWE, and Cybergym, and its price is only 1/3 of Pro. The core reason is: Flash-0731 uses the exact same architecture and parameter scale, but significantly improved its Agentic Capability through Re-Post-Training. V4-Pro-Preview, however, remains stuck on the initial preview version without the same post-training optimization. This phenomenon reveals the critical role of Post-Training in shaping Agentic capabilities, and also shows that scaling is not the only path to improving model capabilities.

1. Background Introduction

On April 24, 2026, DeepSeek officially released the V4 series, including two MoE (Mixture of Experts) models:

Both are released under the MIT open-source license, supporting a 1M Token Context Window and a 384K maximum output length. At launch, the positioning of Pro and Flash was very clear: Pro was the capability ceiling, Flash was the cost-effective choice.

However, on July 31, 2026, DeepSeek released the official V4-Flash-0731 API, completely shattering this landscape. Benchmark data officially released by DeepSeek shows that V4-Flash-0731 comprehensively surpassed V4-Pro-Preview on 9 Agentic benchmarks. Even more surprising is that Flash-0731 and Flash-Preview use the exact same architecture and parameter scale, with the only difference being Re-Post-Training.

References:

2. Architecture Differences

2.1 Parameter Scale Comparison

Specification V4-Pro V4-Flash
Total Parameters 1.6T 284B
Activated Parameters 49B (3.1%) 13B (4.6%)
Number of Experts 256 Experts, 8 activated 64 Experts, 8 activated
Context Window 1M Tokens 1M Tokens
Max Output 384K Tokens 384K Tokens
Precision FP4 + FP8 mixed FP4 + FP8 mixed
Open Source License MIT MIT

2.2 Architectural Commonalities

Both adopt DeepSeek's self-developed Hybrid Attention architecture, combining CSA (Compressed Sparse Attention) and HCA (Heavily Compressed Attention), significantly reducing inference computation and KV Cache usage. Compared to V3.2, V4-Pro's inference FLOPs are reduced by 73%, and KV Cache is reduced by 90%.

2.3 Key Differences

Pro has 4 times the number of experts of Flash (256 vs 64), and 3.8 times the activated parameters (49B vs 13B). From the intuition of Scaling Law, Pro should be superior to Flash in all dimensions. However, on Agentic tasks, Flash-0731 not only did not fall behind but achieved a reversal—this is precisely the core question this document aims to answer.

References:

3. The 0731 Update: The Power of Post-Training

3.1 Core Fact: Same Architecture, Different Results

This is the most critical fact of the entire event: V4-Flash-0731 and V4-Flash-Preview use the exact same architecture and parameter scale, with the only difference being Re-Post-Training.

DeepSeek officially stated in the release notes: "same architecture and size, new post-training". This means that the massive leap in Flash-0731's Agentic capabilities comes entirely from data selection and training strategy optimization in the post-training phase, not from model scale expansion or architectural innovation.

3.2 How Significant is the Effect of Post-Training?

The improvement from Preview to 0731 is astonishing:

Benchmark Flash-Preview Flash-0731 Improvement
DeepSWE 7.3 54.4 +645%
AutomationBench 10.8 25.1 +132%
Cybergym 38.7 76.7 +98%
Toolathlon 49.7 70.3 +41%
NL2Repo 39.4 54.2 +38%
Terminal Bench 2.1 61.8 82.7 +34%
DSBench-FullStack 37.0 68.7 +86%
DSBench-Hard 25.8 59.6 +131%
Agents' Last Exam 15.8 25.2 +59%

This data clearly shows: Post-Training has a far greater impact on Agentic capabilities than model scale itself. Especially DeepSWE improving from 7.3 to 54.4 (+645%), indicating that DeepSeek found key training methods during post-training to significantly enhance Agentic Capability.

3.3 Why is Post-Training So Important for Agentic Tasks?

Agentic tasks (such as Tool Use, Code Generation, Multi-Step Reasoning) are fundamentally different from traditional Language Modeling:

  1. Complex Behavioral Sequences: Agentic tasks require the model to understand the causal relationships of multi-step operations, rather than simple Next Token Prediction
  2. Tool Call Format: Requires precise adherence to API formats and parameter specifications, which demands extremely high format consistency in training data
  3. Error Recovery: Agents need the ability to recover from errors, a capability rarely present in Pre-Training corpora
  4. Feedback Loops: High-quality post-training data can include execution result feedback, allowing the model to learn to adjust behavior based on environmental feedback

DeepSeek likely introduced a large amount of high-quality Agentic Trajectory Data in Flash-0731's post-training, including actual code execution feedback, tool call result verification, etc., thereby achieving a qualitative change in Agentic capabilities at the same parameter scale.

References:

4. Agentic Benchmark Comparison

4.1 Flash-0731 vs Pro-Preview: Comprehensive Surpassing

This is the Agentic Benchmark comparison data officially released by DeepSeek:

Benchmark Flash-0731 Flash-Preview Pro-Preview GLM-5.2 Opus-4.8
Terminal Bench 2.1 82.7 61.8 72.1 81.0 85.0
NL2Repo 54.2 39.4 38.5 48.9 69.7
Cybergym 76.7 38.7 52.7 - 83.1
DeepSWE 54.4 7.3 12.8 46.2 58.0
Toolathlon Verified 70.3 49.7 55.9 59.9 76.2
Agents' Last Exam 25.2 15.8 16.5 23.8 25.7
AutomationBench Public 25.1 10.8 12.8 12.9 27.2
DSBench-FullStack 68.7 37.0 41.8 61.8 71.6
DSBench-Hard 59.6 25.8 31.1 54.5 71.7

Flash-0731 surpasses Pro-Preview on all 9 Agentic Benchmarks. The gaps are most significant especially in DeepSWE (54.4 vs 12.8, 4.3 times higher) and Cybergym (76.7 vs 52.7, 45% higher).

4.2 Verification by Independent Evaluation Agencies

Independent evaluations by Artificial Analysis also confirm this trend: Flash-0731 scores 50 on the Intelligence Index, while V4-Pro scores only 44. In terms of Cost Per Task, Flash is only $0.03, while Pro is $0.05.

4.3 Comparison with Top Industry Models

Flash-0731 has approached or even surpassed GLM-5.2 (Zhipu's high-end model) on most Agentic Benchmarks. Although it still lags behind Claude Opus 4.8, considering that Flash's price is only about 1/50 of Opus 4.8, this cost-effectiveness is unprecedented in the Agentic field.

References:

5. Areas Where Pro Still Dominates

Although Flash-0731 comprehensively surpasses Pro-Preview on Agentic tasks, Pro still maintains a clear advantage in the following areas:

5.1 Knowledge Recall

In the SimpleQA-Verified test, Pro-Preview scores 57.9, while Flash only scores 34.1, a gap of 23.8 percentage points. This indicates that Pro's larger parameter scale has a natural advantage in storing and recalling Factual Knowledge.

5.2 Long Context Retrieval

In the MRCR 1M test, Pro scores 83.5, Flash scores 78.7, a gap of 4.8 percentage points. In large-scale document analysis and precise information localization scenarios, Pro's performance is more reliable.

5.3 Complex Reasoning

On HLE (Humanity's Last Exam), Pro scores 37.7, Flash approximately 34.6, a gap of about 3 percentage points. In scenarios requiring multi-step reasoning and deep analysis, Pro's larger parameter scale still provides a higher ceiling.

5.4 Why Does Pro Have an Advantage in These Areas?

The reason behind this lies in the nature of Scaling Law: knowledge storage and long context processing require the model's "capacity", while Agentic capabilities rely more on "training quality". Pro's 49B activated parameters provide a larger knowledge storage space, but without the same high-quality post-training, this capacity advantage cannot be converted into actual capability on Agentic tasks.

References:

6. Summary

The event of DeepSeek V4-Flash-0731 surpassing V4-Pro-Preview reveals two important laws in AI model capability improvement:

Dimension Key Finding Insight
🧠 The Power of Post-Training Same architecture, same parameters, only through post-training can Agentic capabilities be improved several times over Post-Training is a key link in unleashing model potential, no less important than Pre-Training
⚖️ Mismatch of Scale and Capability Pro with larger parameter scale is surpassed by Flash on Agentic tasks Parameter scale is not the sole determinant of capability; training data quality and strategy are equally critical
🎯 The Specificity of Agentic Capabilities Agentic capabilities depend far more on training data than on parameter scale Agentic capability improvement can be achieved through targeted post-training, without necessarily relying on model scale expansion
💰 Redefining Cost-Effectiveness Flash-0731 achieves higher Agentic performance at 1/3 the price When choosing a model, one needs to combine specific task scenarios and training status, rather than judging solely by parameter scale

Core Conclusion: The root cause of Flash-0731 surpassing Pro-Preview lies in—the time gap in Post-Training. Flash underwent targeted post-training optimization for the 0731 version, while Pro-Preview remains stuck on the initial preview version released in April. When Pro's official version undergoes the same post-training iteration, the gap between the two is likely to widen again. But this event has fully proven: Post-Training is the key to unlocking a model's Agentic capabilities, and its importance may even surpass the expansion of model scale.

Last Updated: 2026-08-11