DeepSeek V4.1 Flash

DeepSeek V4.1 Flash (10 September 2026) is DeepSeek's current main model — a 552B-parameter mixture-of-experts with 8B active parameters for input and 16B for output, a context of up to 1M tokens, and native image input. It replaced V4-Flash and, until a V4.1-Pro arrives, V4-Pro as well.

Best for

• Cheap, high-volume chat, RAG and summarisation
• Long documents and whole codebases in one request
• Reading screenshots, charts and scanned pages
• Self-hosting under a permissive licence

How it compares — and the India angle

• The weights are MIT-licensed — the most permissive licence among current frontier-class open models, which makes it the simplest choice for Indian firms with data-residency rules.
• Calls to deepseek-v4-pro have been routed to V4.1 Flash, at V4.1 Flash prices, since 14 September 2026 — if you still call V4-Pro, you are already using this model.
• DeepSeek charges half price off-peak; schedule batch jobs for those hours.
• The benchmark figures in DeepSeek's release are its own and not yet independently verified.