DeepSeek-V4.1-Flash
DeepSeek · Multimodal model
DeepSeek-V4.1-Flash is the newest DeepSeek model, released September 10, 2026, and the first of a new architecture family with built-in image understanding. It is the main low-cost API model (model name deepseek-flash) and replaced V4-Flash.
Current Open weights
ReleasedSep 10, 2026
Context window1M tokens
Input → outputtext, image → text
API pricedeepseek-flash per 1M tokens: input cache hit $0.003 off-peak / $0.006 peak; input cache miss $0.15 off-peak / $0.3 peak; output $0.6 off-peak / $1.2 peak
Consumer accessDeepSeek chat app and web (chat.deepseek.com), free
Best for
Very cheap, fast default model for coding agents, tool use and image understanding.
Overview
DeepSeek-V4.1-Flash is the newest DeepSeek model, released September 10, 2026, and the first of a new architecture family with built-in image understanding. It is the main low-cost API model (model name deepseek-flash) and replaced V4-Flash.
Strengths and limits
- Extremely low API price with 1M context and up to 384K output
- Reported scores ahead of DeepSeek-V4-Pro on many agent benchmarks
- MIT-licensed open weights with native vision input
- Very large (552B backbone) to self-host despite the 'Flash' name
- Benchmarks are self-reported by DeepSeek at max reasoning effort
Published benchmark scores
GPQA Diamond (Pass@1)90.9
Source: huggingface.co
Terminal-Bench 2.1 (Pass@1)90.6
Source: huggingface.co
DeepSWE v1.1 (Resolved)74.2
Source: huggingface.co
Codeforces (Rating)3471
Source: huggingface.co
Scores are as published by the source; settings (reasoning effort, tools, attempts) differ between labs, so compare with care. How to read benchmarks →