DeepSeek releases V4-Flash-0731 with major performance gains

Jul 31, 2026

DeepSeek released V4-Flash-0731 on July 31, 2026, a major upgrade to its AI model with significant gains in agentic tasks. The 284B-parameter model now scores 50 on the Artificial Analysis index, nearly matching OpenAI's GPT-5.6 Luna at a 60% lower cost. The update, achieved via post-training, is available through a public beta API and as MIT-licensed weights on Hugging Face.

V4-Flash-0731 release details

▪The model's weights are available on Hugging Face under an MIT license, allowing for on-premise commercial deployment.
▪DeepSeek released the V4-Flash-0731 model on Hugging Face and moved its V4-Flash API into public beta on July 31, 2026.
▪The 0731 release is an update to the preview version from April 2026, with gains from re-post-training rather than architectural changes.

API pricing structure

▪Per-task costs for V4-Flash-0731 are about 60% lower than for OpenAI's GPT-5.6 Luna.
▪The API offers a 98% cache discount, with cache hits costing $0.0028 per 1M input tokens.
▪The V4-Flash API is priced at $0.14 per 1M input tokens and $0.28 per 1M output tokens.

Self-hosting requirements

▪A full-precision version of the model can be served on a single 4xGB300 node using DeepSeek's vLLM example.
▪Self-hosting the model requires significant memory, with a 3-bit quantized build needing around 110 GB of combined RAM and VRAM.

Model architecture specifications

▪V4-Flash is a 284-billion-parameter Mixture-of-Experts (MoE) model that activates 13 billion parameters per token.
▪The model features a 1-million-token context window.
▪The checkpoint includes the DSpark speculative decoding module, bringing the total parameter count in the Hugging Face repo to 304B.

Benchmark performance results

▪DeepSeek's reported benchmarks, including an 82.7 on Terminal Bench 2.1, were run using an unreleased "minimal mode" of its testing harness.
▪The model's score on the GDPval benchmark, which tests complex office work, increased from 1,189 to 1,559 Elo points.
▪On the Artificial Analysis Intelligence Index, V4-Flash-0731 scored 50 points, just one point behind OpenAI's GPT-5.6 Luna.

Serving configuration options

▪The DSpark speculative decoding module can be enabled in vLLM and is reported to speed up per-user generation by 60-85%.
▪For agentic tasks, DeepSeek recommends setting temperature to 1.0 and top_p to 0.95.