DeepSeek-V4-Flash-0731: The Smaller Model That Beat Pro on Agentic Benchmarks

On July 31, 2026, DeepSeek released the official version of DeepSeek-V4-Flash-0731, superseding the preview. This is not a routine update — it’s a situation where the smaller model (284B total, 13B activated parameters) consistently beats its bigger sibling DeepSeek-V4-Pro (1.6T total, 49B activated) on agentic benchmarks. The engineering equivalent of winning a race with a 1.2-liter engine instead of a 5.0.

Architecture: What Changed

V4-Flash-0731 keeps the V4-series architecture, with three key upgrades over V3.2:

• Hybrid attention (CSA + HCA) — Compressed Sparse Attention and Heavily Compressed Attention cut inference cost at 1M-token context. V4-Pro needs only 27% of FLOPs and 10% of KV cache compared with V3.2 at full context.
• Manifold-Constrained Hyper-Connections (mHC) — strengthens residual connections, improving signal propagation stability.
• Muon optimizer — faster convergence, greater training stability.

The model was pre-trained on 32T+ tokens, with a two-stage post-training pipeline: independent cultivation of domain experts (SFT + RL with GRPO), then consolidation via on-policy distillation.

Benchmarks: Flash vs Pro vs the World

Flash-0731 beats Pro on every agentic benchmark. The biggest gap: DeepSWE (54.4 vs 12.8 — over 4x) and Cybergym (76.7 vs 52.7). The jump over its own preview version is even more dramatic — 2x to 7x depending on the test.

What This Means for Engineers

This confirms the trend we’ve been talking about for months: architectural efficiency beats raw scale. 13B activated parameters + good attention design + solid post-training = better results than 49B activated without those improvements.

Practical consequences: lower inference costs (smaller model = less memory and FLOPs), faster responses, and — most importantly — MIT license. Plus a speculative decoding module (same as the DSpark version) that speeds up generation without quality loss.

This content was generated using artificial intelligence and reviewed by a human editor.

Sources:
– HuggingFace: DeepSeek-V4-Flash (official card)
– arXiv: DeepSeek-V4 Technical Report
– HuggingFace: DeepSeek-V4-Flash-0731
– Polish version: [PL] DeepSeek-V4-Flash-0731