DeepSeek V4 Flash 0731: open-weights model scores 35 on Intelligence Index

Artificial Analysis has published its analysis of DeepSeek V4 Flash 0731, the reasoning variant of DeepSeek's open-weights model evaluated at maximum effort. It scores 35 on the firm's Intelligence Index, well above the median of 18 recorded for comparable open-weight models, and receives 4 out of 4 units for both intelligence and speed against 2 out of 4 for cost. The model was released on July 31, 2026 and its weights are published for download.

Intelligence and token use

The Intelligence Index is a composite benchmark that covers reasoning, knowledge, mathematics and coding. At 35 points the model sits in the upper range of its size class, defined as open-weight models with more than 150 billion parameters. The evaluation also tracks how many tokens a model needs to reach its score: DeepSeek V4 Flash 0731 generated 240 million output tokens, against a median of 130 million for similar models, which Artificial Analysis classifies as very verbose.

Speed and price

Measured through DeepSeek's API, the model produces 213.2 tokens per second — well above the 76.2 tokens per second median for comparable open-weight models — and reaches its first token in 1.17 seconds against a median of 2.30 seconds. Pricing is listed at $0.44 per million input tokens and $1.32 per million output tokens. On a blended 7:2:1 cache-hit, input and output ratio the effective rate comes to $0.23 per million tokens, with a reported cache discount of 97%. One Intelligence Index task costs a weighted $0.22, and the full evaluation run totalled $474.19.

Architecture and availability

DeepSeek V4 Flash 0731 is a mixture-of-experts model with 284 billion total parameters and 13 billion active per token. It accepts text and returns text only, with no image input, and its context window reaches one million tokens — roughly 1,500 A4 pages. The weights are available on Hugging Face under the MIT licence, which permits commercial use, and the model can be called through 17 API providers.