跪拜 Guibai
← All articles
Artificial Intelligence

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

By 郑恩赐 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A model one-sixth the size beating a flagship on agentic benchmarks upends the assumption that bigger is always better for tool-using AI. Teams building autonomous coding or automation agents can now get state-of-the-art performance at a fraction of the inference cost — provided the model has been post-trained on high-quality agent trajectories.

Summary

DeepSeek V4-Flash-0731, a 284B-parameter MoE model with 13B active parameters, scores higher than the 1.6T-parameter V4-Pro-Preview across all nine agentic benchmarks — including a 645% jump on DeepSWE. The two models share identical architecture and parameter counts with their preview versions; the only difference is a round of re-post-training on Flash.

Pro still leads on knowledge recall, long-context retrieval, and complex reasoning benchmarks like HLE, where raw parameter capacity matters more. But on agentic tasks — tool use, multi-step coding, error recovery — post-training quality swamps model size. Independent evaluator Artificial Analysis pegs Flash-0731 at an Intelligence Index of 50 versus Pro's 44, with a per-task cost of $0.03 against $0.05.

The result is a timing artifact: Pro-Preview hasn't received the same post-training update yet. Once it does, the gap should widen again in Pro's favor. The takeaway is that agentic capability is a training-data problem, not a scaling problem.

Takeaways
V4-Flash-0731 uses the same architecture and parameter count as the earlier Flash-Preview; only the post-training changed.
Flash-0731 beats V4-Pro-Preview on all nine agentic benchmarks, with DeepSWE jumping from 7.3 to 54.4 (+645%).
Flash-0731 costs $0.03 per task versus Pro's $0.05, while scoring 50 vs. 44 on Artificial Analysis's Intelligence Index.
Pro still wins on factual knowledge (SimpleQA 57.9 vs. 34.1), long-context retrieval (MRCR 1M 83.5 vs. 78.7), and complex reasoning (HLE 37.7 vs. ~34.6).
Agentic capability depends far more on post-training data quality — tool-call formats, execution feedback, error-recovery trajectories — than on parameter count.
Pro-Preview is still the April preview version; once it gets the same post-training treatment, the performance order will likely flip back.
Conclusions

The 645% DeepSWE gain from post-training alone suggests that pre-training on code leaves enormous latent agentic capability that standard fine-tuning never surfaces.

Calling this a 'Flash beats Pro' story is misleading — it's a 'post-trained model beats stale preview' story. The real comparison will come when Pro receives equivalent treatment.

Agentic benchmarks are measuring something fundamentally different from knowledge or reasoning benchmarks, and the industry's habit of ranking models on a single leaderboard obscures this split.

At $0.03 per task with Opus 4.8-competitive scores, Flash-0731 resets the price-performance curve for agent hosting. The economic pressure on proprietary labs to match this will be significant.

Concepts & terms
Post-Training
A phase after initial pre-training where a model is fine-tuned on curated data — here, agent trajectories with execution feedback, tool-call formats, and error-recovery sequences — to shape specific behaviors without changing the model architecture or parameter count.
Agentic Benchmarks
Evaluations that measure a model's ability to act autonomously: using tools, writing and executing code across multiple steps, recovering from errors, and completing multi-stage tasks (e.g., DeepSWE, Cybergym, Terminal Bench). They test behavioral sequences rather than static knowledge.
MoE (Mixture of Experts)
A model architecture where only a subset of parameters ('experts') are activated for any given input. DeepSeek V4-Pro has 256 experts with 8 active (49B activated out of 1.6T total); V4-Flash has 64 experts with 8 active (13B activated out of 284B total).
Source: juejin.cn ↗ Google Translate ↗ Backup ↗