DeepSeek V4.1 Flash

MIT open weights and hosted DeepSeek API access as deepseek-flash.

Publisher-reported capabilities

• A new causal encoder-decoder mixture-of-experts architecture with native image and text input, text output, and a 1M-token context window.
• The API supports thinking and non-thinking modes, tool calls, and Responses and Anthropic-compatible interfaces.

Limitations and availability caveats

• Peak and off-peak API prices differ; the stored prices use peak rates.
• The reported 552B parameter count describes the backbone; the model also has 196B of Engram conditional memory.

Official release and model references

• api-docs.deepseek.com/news/news260910/
• api-docs.deepseek.com/quick_start/pricing
• huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash

Deepseek V4.1 Flash has 22 of 39 tracked IQ benchmark inputs across 7 of 7 dimensions. Raw benchmark tables contain source-backed results only. Derived IQ combines all six dimensions using the shared scoring methodology, including estimates where direct coverage is incomplete. Configured predecessor references: Deepseek V4 Flash. A reference is an estimation input, not evidence that this model has the predecessor's measured results.

Among dimensions with direct benchmark coverage, this profile's highest estimated scores are Programmatic Reasoning (IQ 135) and Mathematical Reasoning (IQ 131). These dimension estimates can include statistical adjustments or predecessor priors; they are not raw benchmark percentages or claims of real-world task success.

Benchmark versions, evaluation harnesses, and coverage affect comparisons; a missing result is not a score of zero. See the scoring methodology for estimation and uncertainty limits.