DeepSeek-V4.1-Flash
Language modelOpen weightsCard summarizedThe smallest model in our new architecture family, with native visual understanding
Active (latest)● Re-checked 12 Oct 2026 · 78 facts · 0 corrections · 1 unconfirmedSource: DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression (51 pp)
Use it via API
Where to call it
- DeepSeek API
deepseek-flash - Self-host (weights)
deepseek-ai/DeepSeek-V4.1-Flash
Good for
- Agentic coding: 90.6% on Terminal-Bench 2.1 and 74.2% on DeepSWE v1.1
- Everyday coding tasks and white-collar workflows at low serving cost
- Competitive programming and math (Codeforces 3471, MathArena Apex 65.6%)
Overview
V4.1 Flash
The smallest model in our new architecture family, with native visual understanding
- Cost
- $0.30 in
- Context
- 1M
- Input / output
- $0.30 / $1.20
- Max output
- 384K
- Knows up to
- —
- Released
- 10 Sep 2026
A multimodal Mixture-of-Experts model with 552B backbone parameters and a 1M-token context. A causal encoder-decoder design activates 8B parameters per token during prefill and 16B during decode, and KV cache compression cuts the global KV cache to about 1/4 of DeepSeek-V4-Flash. DeepSeek reports it beats DeepSeek-V4-Pro on most agentic benchmarks. p.1
deepseek-flash- Input / output per 1M
- $0.30 / $1.20
- Context window
- 1M
- Max output
- 384K
- Knows up to
- —
- Reasoning
- Configurable (thinking on by default; low / high / max)
- Weights
- Open · MIT
About this card. A model card is the maker's own account of a model: what it's for, how it was trained and tested, and where it falls short. This page condenses DeepSeek-AI's DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression. Every claim carries its source page.