DeepSeek V4 Flash

DeepSeek V4 Flash 0731 is the official release of DeepSeek's small agentic model, shipped July 31, 2026 with MIT-licensed open weights the same day. It keeps the 284B-total / 13B-active MoE architecture from the April preview but was re-post-trained specifically for agentic performance, with the DSpark speculative decoding module attached and native Responses API plus Codex support added. On DeepSeek's published benchmarks the official Flash build now outruns the much larger V4-Pro preview, making it one of the best cost-per-task models in the frontier agentic coding space.

Core Features

• Re-post-trained for agents: Same architecture as the preview, but the post-training pass focuses on tool use, multi-step tasks, and keeping full reasoning history in context across every tool-calling round.
• DSpark speculative decoding: Built-in draft module speeds inference; enable with a single --speculative-config flag on vLLM or DSPARK on SGLang.
• Three reasoning efforts: low, high, and max control deliberation, with max pushing full problem decomposition and edge-case testing; non-thinking mode is also supported.
• Codex-ready API: Native OpenAI Responses API support and explicit adaptation for Codex, plus an Anthropic-compatible /anthropic endpoint.
• Massive context: 1M-token context with up to 384K output tokens, keeping long agent sessions in one window.

Model Specifications

Pricing

A 2x peak-hours surcharge (09:00-12:00 and 14:00-18:00 Beijing time) was announced but not yet effective at release.

Benchmark Highlights

• Terminal Bench 2.1: 82.7, more than twenty points above the preview build.
• Agentic benchmarks: Outperforms the DeepSeek-V4-Pro preview on every agentic benchmark DeepSeek publishes, despite being roughly 12x smaller in active parameters.
• Ecosystem: The model antirez's DwarfStar 4 (ds4) native C inference engine was built around.

Use Cases

• Cost-sensitive agent pipelines: Frontier-adjacent agentic performance at $0.14/$0.28 per 1M tokens.
• Codex-compatible workflows: Teams already on OpenAI-shaped APIs can switch the model name and keep their stack.
• Local agent deployment: MIT weights run on commodity hardware with vLLM, SGLang, or llama.cpp-class engines.
• High-concurrency services: Up to 2,500 concurrent requests on the API.

Advantages

1. Price-performance king: 13B active parameters that outscore a 49B-active flagship on agentic benchmarks.
2. Open and practical: MIT weights with DSpark and a huge context make it the strongest open agentic model per dollar.
3. Ecosystem fit: Native Codex adaptation and Anthropic-compatible endpoints remove integration friction.

Tips

1. Enable DSpark: Speculative decoding is the headline speed feature; use the documented flag on your serving stack.
2. Raise output limits for max effort: The max reasoning level benefits from the 384K output ceiling.
3. Exploit the 98% cache discount: Long shared prefixes cost nearly nothing on cache hits, which changes how you structure agent prompts.

Conclusion

DeepSeek V4 Flash 0731 is the definitive small agentic model of the summer: open weights, 1M context, and a cost profile that makes frontier-adjacent agent coding affordable at scale. If your team builds agents on a budget, it is the strongest starting point.

Alternatives

• DeepSeek V4 Pro 0813: The 1.6T flagship for maximum capability when cost is secondary.
• GLM-5.3: The new open-weight coding leader at slightly higher cost.
• Kimi K3: Open-weight 2.8T model with comparable agentic scores.