DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains
DeepSeek published DeepSeek-V4-Flash-0731 on Hugging Face and moved the official V4-Flash API into public beta on July 31, 2026. Architecture and size unchanged from the April preview (284B MoE / 13B active, 1M context); gains come from re-post-training. Checkpoint ships with DSpark speculative decoding module attached.
API: deepseek-v4-flash now natively supports the Responses API format and is adapted for Codex. V4-Pro API and app/web models were not updated.
Pricing (DeepSeek, as reported)
- 0.0028 cache hit, $0.28 / 1M output
- ~1/3 of deepseek-v4-pro output pricing ($0.87)
- 2,500 concurrency limit
Weights: MIT-licensed and ungated. Self-hosting needs substantial memory (Unsloth: ~103 GB at 3-bit / ~162 GB 8-bit; DeepSeek vLLM example on 4×GB300).
Selected vendor-reported agent benchmarks (V4-Flash-0731)
| Benchmark | 0731 | Flash Preview | V4-Pro Preview |
|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 |
| Toolathlon-Verified | 70.3 | 49.7 | 55.9 |
| Cybergym | 76.7 | 38.7 | 52.7 |
| DeepSWE | 54.4 | 7.3 | 12.8 |
Caveats: Code Agent tasks used unreleased DeepSeek Harness minimal mode; some sets internal; scores harness-sensitive.
HF model: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731