DeepSeek V4-Flash is the cheap, high-throughput half of DeepSeek's V4 model family, a 1M-context model aimed at developers running long agent loops and coding agents where a single task can trigger hundreds of model calls. The V4-Flash-0731 checkpoint landed on Hugging Face and the API entered public beta on 31 July 2026. It is a 284B-parameter mixture-of-experts model that activates only 13B parameters per token, released under an ungated MIT licence.
Core Features
- 1M-token context window with up to 384K tokens of output in one response
- Thinking and non-thinking modes switch per request; the retiring deepseek-chat and deepseek-reasoner names map onto those two modes
- Tool calling, JSON output, FIM completion and chat prefix completion for agent harnesses
- Served in both OpenAI and Anthropic API formats at api.deepseek.com, with Responses API support and Codex compatibility
- DSpark speculative decoding module bundled into the weights, enabled in vLLM with a single flag
- Concurrency ceiling of 2,500 requests, five times what V4-Pro allows
Use Cases
- Coding agents and CI pipelines where token spend compounds across long multi-step runs
- Whole-codebase or long-document analysis that actually needs the 1M context
- Self-hosted deployment for teams that cannot send source code to a third-party API
Pricing
API rates are $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output tokens, roughly a third of what V4-Pro charges. Model weights are free to download and run on your own hardware.
Our Take
Best for high-volume agent and coding workloads where call count matters more than peak reasoning; keep V4-Pro or Claude for the few hardest steps. The trade-off worth pricing in: DeepSeek ran its agent benchmarks at maximum reasoning effort, so real-world token spend can climb well above the headline rate. More engines at aifreetool.site/tool-category/ai-engine-model/
