DeepSeek V4 Flash 0731 (Alibaba Cloud)

DeepSeek V4 Flash 0731 is the official release of the DeepSeek V4 Flash line, replacing its preview with a version designed around improved agentic work. Its architecture follows the DeepSeek V4 Flash DSpark design and includes an attached speculative decoding module. That construction supports the model’s focus on efficient model execution while retaining a model family lineage centered on practical agent workflows rather than only conversational text generation.

The reported evaluation results position this release as a strong option for coding and tool-driven agents. It scores 82.7 on Terminal Bench 2.1 and 54.2 on NL2Repo, exceeding the cited DeepSeek V4 Pro Preview results of 72.1 and 38.5 despite a much smaller activated parameter count. It also records 76.7 on Cybergym, 54.4 on DeepSWE, and 70.3 on Toolathlon-Verified. These results make it a practical fit for software engineering, repository work, and multi-step tool use where agent capability and execution efficiency matter.

LLM Gatewayalibaba/deepseek-v4-flash-0731deepseek-flash

Open provider page Provider docs

Quick Info

Provider: LLM Gateway
Model key: alibaba/deepseek-v4-flash-0731
Release date: Jul 31, 2026
Last updated: Jul 31, 2026
Knowledge cutoff: 2025-05
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.44
Output token cost: $1.32

Limits

Output tokens: 393,216 tokens
Context window: 1,000,000 tokens

Transparent token rates

Compare DeepSeek V4 Flash 0731 (Alibaba Cloud) pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about DeepSeek V4 Flash 0731 (Alibaba Cloud)

LLM Gateway

CoveragePreview

DeepSeek released DeepSeek-V4-Flash-0731 on July 31, 2026, and began an API public beta. According to translated XenoSpectrum coverage, the update added native Responses API support, Codex configuration, and low, high, and max reasoning-intensity settings. The report attributes the release to DeepSeek, not the gateway serving the model. The reported release retained the Preview architecture at 280 billion total parameters with 13 billion active per token under a mixture-of-experts design. XenoSpectrum says the model supports a 1 million-token context, up to 384,000 output tokens, and MIT-licensed weights for self-hosting with vLLM or SGLang, while large-scale pretraining was not rerun.

Sulat also reports on AI and tech beyond the catalog.Read news.sulat.com