DeepSeek-V4.1-Flash

Maker

DeepSeek

Released

10 Sep 2026

Open weights

Yes, you can download it and run it yourself

License

MIT: free for almost any use

Memory (context window)

1 million tokens (how much it can read at once)1,000,000 tokens exactly

Size

763B total (Hugging Face count) parametersB = billion, T = trillion. Bigger usually needs a stronger computer.

Price

$0.3 for every million tokens you send, $1.2 for every million tokens it writesA token is a small piece of a word · checked 17 Oct 2026 · price source

Official page

Official page for DeepSeek-V4.1-Flash (huggingface.co)

• All-roundArtificial Analysis Intelligence Index

  39 index points Tied #10 of 17 models scored independent

  One combined score from many independent tests of how smart a model is.

  setting: max · as of 17 Oct 2026 · source

• Chat & writingLMArena Text

  1475 Arena score #8 of 15 models scored independent

  When real people compare two AI answers side by side without knowing which AI wrote them, which one do they pick more often?

  setting: Max · as of 15 Oct 2026 · source

• Build websites & appsLMArena WebDev

  1618 Arena score #9 of 18 models scored independent

  When people ask two AIs to build the same website or app and compare the results, which one do they pick?

  setting: Max · as of 17 Oct 2026 · source

• CodingLiveBench 2.0 Coding

  85.4 score out of 100 #6 of 15 models scored independent

  Fresh coding tasks with tests that check the answer, changed often so models cannot memorize them.

  setting: highest supported effort · as of 14 Oct 2026 · source

• ReasoningLiveBench 2.0 Reasoning

  86.7 score out of 100 #6 of 15 models scored independent

  Fresh logic and puzzle questions with one right answer, changed often so models cannot memorize them.

  setting: highest supported effort · as of 14 Oct 2026 · source

• Science questionsGPQA Diamond

  90.9% #3 of 6 models scored maker's own number

  Very hard science questions in biology, chemistry and physics, written by PhD experts.

  as of 17 Oct 2026 · source

• Hardest examHumanity's Last Exam (no tools)

  36.8% #4 of 6 models scored maker's own number

  2,500 expert questions written to be too hard for AI, answered without web search or code.

  as of 17 Oct 2026 · source

• New puzzlesARC-AGI-2

  72.9% #5 of 11 models scored independent

  New picture puzzles a model has never seen, so it has to reason instead of remember. Harder for AI than for people.

  setting: Max, $0.129 per task · as of 18 Oct 2026 · source

• Most used (this week)OpenRouter usage (this week)

  36.8 trillion tokens this week #1 of 4 models scored independent

  How much developers actually use each model through OpenRouter, a popular service that connects apps to many AI models.

  setting: week to 16 Oct · as of 16 Oct 2026 · source

• SpeedOutput speed

  176 tokens per second #4 of 16 models scored independent

  How many tokens (pieces of words) per second the model writes once it starts answering. Higher is faster.

  setting: max · as of 17 Oct 2026 · source

• MetaMuse Spark OpenOpen weights. License: Muse model license: free to use with conditions39 index pointsTied #10 of 17 models scored
• OpenAIGPT-6 Luna38 index points#12 of 17 models scored
• AnthropicClaude Haiku 5.543 index points#9 of 17 models scored

No stories about DeepSeek-V4.1-Flash on ITG Now yet.

More from DeepSeek

• DeepSeek · 17 OctDeepSeek says free model update tops a coding test
• DeepSeek · 17 OctDeepSeek opens search API at $0.75 per 1,000

• DeepSeek Harness: DeepSeek's Free Open-Source AI AgentDeepSeek Harness is DeepSeek's free, open-source AI agent. Every feature is a plugin you can swap or build, and it runs in your browser with one command.