DeepSeek V4.1 Flash
DeepSeek · Vision-language · MoE 552B (A8-16B) · 1000k context · Released 10 September 2026
DeepSeek V4.1 Flash was published on 10 September 2026 with the weights on Hugging Face under the MIT licence. It departs from the decoder-only shape almost every current model uses, adopting a Causal Encoder-Decoder design of a 20-layer causal encoder followed by a 20-layer decoder, and it activates roughly 8 billion parameters per token while reading the prompt and roughly 16 billion while generating, a deliberate asymmetry that makes reading long inputs cheap. It takes images and text and produces text, carries a 1-million-token context, and was trained on a reported 45 trillion tokens of mixed text and images. At this scale it is a GPU-server model rather than a consumer-card one, and local runtime support for the encoder-decoder architecture should be confirmed before relying on it.
Strengths
- MIT-licensed, with no revenue threshold and no territory restriction
- Multimodal input (images and text) with a 1-million-token context
- Very sparse: only about 8 to 16 billion of 552 billion parameters compute per token
- Efficient to serve at high volume relative to its total size
Weaknesses
- Server-class, needing a multi-GPU server rather than a workstation card
- An unusual encoder-decoder architecture, so local runtime support may lag
- Independent benchmark results are limited at this stage
- Reads images but does not generate them
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| FP8 (as released) | ~552GB | The released format, roughly 552GB across many shards, close to one byte per parameter. Multi-GPU server territory, and actual needs rise with context and concurrency. Figure is approximate. |
| BF16 | ~1100GB | Roughly 1.1TB at full precision. A community 4-bit quantisation would cut the released size to around 280GB but remains server-class. Figures are approximate. |
What you'd need to run this
Licence
MIT — read the licence
DeepSeek V4.1 Flash: common questions
- What hardware do I need to run DeepSeek V4.1 Flash?
- At its most compressed (FP8 (as released)) it needs roughly 552GB of VRAM. VRAM figures are approximate and depend on context length and settings.
- Is DeepSeek V4.1 Flash free for commercial use?
- Yes. DeepSeek V4.1 Flash is licensed under MIT, which permits commercial use with no meaningful conditions.
- Can I run DeepSeek V4.1 Flash on Apple Silicon?
- It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.
- What is DeepSeek V4.1 Flash's context window?
- DeepSeek V4.1 Flash has a context window of 1,000,000 tokens, about 1000k.
Availability
Recommended for
- High-volume multimodal understanding kept in-house under a permissive licence
- Server-class self-hosting where data must stay on your infrastructure
- Teams comparing a hosted multimodal API against openly licensed self-hosting
Related models
Related guides
Glossary
Catalogue entry last verified 24 September 2026. Specifications change; verify anything you are about to spend money on.