DeepSeek V4 Flash: Agent Power at Flash Speed

284B total, 13B active, 1M context. The official V4 Flash release beats V4 Pro Preview on agent benchmarks — at one-third the cost. MIT open weights on Hugging Face.

See Benchmarks → API Pricing
284B
Total Parameters
13B
Active Parameters
1M
Context Window
$0.14/M
Input Token Price

What Is DeepSeek V4 Flash?

DeepSeek V4 Flash is a 284B-parameter Mixture-of-Experts model with 13B active parameters per token, officially released on July 31, 2026. It is the smaller, faster sibling of DeepSeek V4 Pro (1.6T/49B), yet its official 0731 build exceeds V4 Pro Preview on every agent benchmark — a result DeepSeek attributes entirely to improved post-training, not architectural changes.

The model offers a 1M-token context window, MIT open-weight license on Hugging Face, and API pricing at $0.14 per million input tokens — making it roughly three times cheaper than V4 Pro and over twenty times cheaper than Claude Sonnet 4.6. Community benchmarks show 400 tokens per second on rental hardware at about $10 per hour, making self-hosting practical even for small teams.

V4 Flash's sweet spot is agentic workloads: tool use, multi-step coding, terminal automation, and MCP orchestration. It scores 49.1% on Terminal-Bench 2.0, 64% on MCP Atlas, and 57.8% on Claw-Eval — numbers that put a 13B-active model in territory previously reserved for models four to five times its size. This site is an unofficial guide covering benchmarks, pricing, and comparisons.

Why Developers Are Switching to V4 Flash

Six reasons the smallest V4 model is getting the most attention.

Agent power in a flash model

The 0731 official release was only re-post-trained — same 284B architecture — yet agent benchmarks now far exceed V4 Pro Preview, the 1.6T model in its own family.

$0.14 per million input tokens

Three times cheaper than V4 Pro and over twenty times cheaper than Claude Sonnet. For high-volume agent workloads, this makes experiments affordable.

1M token context window

Feed entire codebases, long documents, or hours of transcribed audio into a single request — the same context length as V4 Pro at a fraction of the cost.

400 tokens per second self-hosted

Community benchmarks show 400 tps on rented hardware at about $10 per hour. The 13B active parameter count makes V4 Flash practical to self-host.

MIT open weights

Full weights on Hugging Face under MIT license. Download, modify, fine-tune, and deploy commercially with zero restrictions.

Drop-in API replacement

Set model to deepseek-v4-flash in any OpenAI-compatible client. The API calling method is unchanged from the preview — upgrade with one config change.

Explore the Guides

Benchmarks →

Agent, coding, and reasoning scores — with verification status for each. See why V4 Flash beats models 4x its size.

API Pricing →

$0.14/M input, $0.28/M output. Side-by-side with V4 Pro, Qwen, GPT, and Claude. Plus self-hosting cost math.

vs Other Models →

V4 Flash vs V4 Pro, Qwen3.7 Flash, and GPT-5.5 Mini — parameters, benchmarks, pricing, and which to pick.

DeepSeek V4 Flash FAQ

What is DeepSeek V4 Flash?
It is a 284B-parameter Mixture-of-Experts model with 13B active parameters per token, released officially on July 31, 2026. It offers a 1M-token context window, MIT license, and API pricing at $0.14 per million input tokens — making it one of the cheapest frontier-class models available.
How is V4 Flash different from V4 Pro?
V4 Pro has 1.6T total parameters with 49B active — roughly four times larger. Yet the official V4 Flash 0731 release exceeds V4 Pro Preview on agent benchmarks despite being much smaller and three times cheaper per token.
Is DeepSeek V4 Flash open source?
Yes. Full weights are on Hugging Face under MIT license. You can download, modify, fine-tune, and deploy commercially. The API is also available through DeepSeek's platform in public beta.
What changed in the 0731 release?
DeepSeek says V4 Flash 0731 was only re-post-trained — same architecture, same parameter count. The improvement is entirely in post-training, specifically targeting agent capabilities, which now far exceed the larger V4 Pro Preview.
How much does V4 Flash cost?
$0.14 per million input tokens and $0.28 per million output tokens through the DeepSeek API. Self-hosting on rented GPUs runs about $10 per hour at 400 tokens per second.
Can I self-host DeepSeek V4 Flash?
Yes, and it is practical — 13B active parameters means the model runs efficiently on modest GPU setups. Community reports show 400 tokens per second on standard rental hardware at approximately $10 per hour.
What is V4 Flash best at?
Agent tasks, coding, and tool use. The 0731 release specifically targets agentic workloads — Terminal-Bench 49.1%, MCP Atlas 64%, Claw-Eval 57.8%. It also performs well on coding and reasoning benchmarks.
How does V4 Flash compare to GPT and Claude?
It is over 20x cheaper than Claude Sonnet 4.6 per token and 4x cheaper than GPT-5.5 Mini. On agent benchmarks it is competitive with far more expensive closed models, though top-tier closed models still lead on some general reasoning tasks.

Get Started Today

Set your model to deepseek-v4-flash in any OpenAI-compatible client. Or download the MIT weights and self-host at 400 tps.

See Benchmarks Weights on HF