DeepSeek Upgrades V4-Flash with 0731 Checkpoint: Major Gains from Re-Post-Training
DeepSeek has published DeepSeek-V4-Flash-0731 on Hugging Face and moved the official V4-Flash API into public beta. The architecture and 284B-parameter size remain unchanged — the improvements come entirely from re-post-training. The checkpoint ships with the DSpark speculative decoding module attached, bringing the total to 304B parameters. The API now natively supports the Responses API format and is adapted for Codex.
On DeepSeek-reported benchmarks, the 0731 version dramatically outperforms both the V4-Flash preview and V4-Pro on agentic and coding tasks: 82.7% on TerminalBench 2.1 (up from 61.8), 54.4% on DeepSWE (up from 7.3), and 76.7% on Cybergym (up from 38.7). Self-hosting requires roughly 110 GB combined RAM+VRAM at 3-bit quantization or a 4×GB300 node at full precision. The model is MIT-licensed and ungated, with API pricing at $0.14/1M input tokens (cache miss) and $0.28/1M output tokens — roughly a third of V4-Pro output pricing.
Related: DeepSeek Launches V4-Flash API with Agentic Upgrades