DeepSeek V4.1 Flash
A 552B mixture-of-experts model with a tiny KV cache, native image input, 1M context and the MIT license.
In short. DeepSeek V4.1 Flash is a very capable open model whose main trick is saving memory when working with long inputs, which makes long coding and agent sessions much cheaper to run. It also understands images, and the MIT license lets you use it however you like. It still needs a multi-GPU server.
Summary
DeepSeek V4.1 Flash was released on 10 September 2026 as the smallest model of DeepSeek's new architecture family, and it adds native image understanding. On 17 October 2026 DeepSeek replaced the weights with an updated checkpoint, DeepSeek-V4.1-Flash-1017, which it calls a minor update with better reasoning, coding, agent and visual results. On the DeepSeek API both are reached with the name deepseek-flash.
It is a mixture-of-experts model with 552 billion backbone parameters. Its Causal Encoder-Decoder design splits 40 layers into a 20-layer encoder and a 20-layer decoder, so only about 8 billion parameters are active per token while reading the prompt and about 16 billion while generating. That suits agent workloads, which read far more text than they write. The context window is 1 million tokens.
The headline feature is KV-cache compression. By reusing attention indices across layers and storing the cache in FP4, the technical report puts the always-in-GPU-memory cache at 890 bytes per token, about a quarter of DeepSeek V4 Flash, with about one eighth of the persistent cache. The model was pretrained on 45 trillion multimodal tokens. The checkpoint is about 510 GB in mixed FP8 and FP4 form, and both the code and weights use the MIT license.
DeepSeek's benchmark claims, for example Terminal-Bench 4.0 rising from 31.2 to 42.3 in the 1017 update, are self-reported, and DeepSeek says part of the gain comes from tuning for specific agent harnesses.
Good for
- Long-context agent and coding workloads where cache memory dominates cost
- Products that need the MIT license with no revenue limits
- Reading images together with code and documents
- Teams already running vLLM or SGLang on multi-GPU nodes
Watch out for
- The checkpoint is about 510 GB, so it needs a multi-GPU node and is not a laptop model
- It writes text only
- Benchmark gains are self-reported and partly come from harness-specific tuning
- Old API names such as deepseek-v4-flash are now served by the new model, so pinned tests may change
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the V4.1-Flash variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| V4.1-Flash-1017 FP4 | 552B 16B active | 1.27 TB | 635 GB | 562 GB | GPU server |
| V4.1-Flash FP4 | 552B 16B active | 1.27 TB | 635 GB | 561 GB | GPU server |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- GLM-5.3 vs DeepSeek V4.1 Flash: open models for coding agents compared
- Best open-source LLMs in 2026
- Meituan's LongCat-2.5: 1.6T parameters, vision and the plain MIT license
- Giant open models now ship in 4-bit: what it means for memory planning
Frequently asked questions
- What is DeepSeek V4.1 Flash?
- DeepSeek V4.1 Flash is a very capable open model whose main trick is saving memory when working with long inputs, which makes long coding and agent sessions much cheaper to run. It also understands images, and the MIT license lets you use it however you like. It still needs a multi-GPU server.
- Can I use DeepSeek V4.1 Flash commercially?
- Yes. DeepSeek V4.1 Flash is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is DeepSeek V4.1 Flash open source?
- DeepSeek V4.1 Flash is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run DeepSeek V4.1 Flash on my own computer?
- Not on a desktop. The smallest variant, V4.1-Flash, needs about 561 GB at 4-bit precision, which means a GPU server with several data-center cards.
- When was DeepSeek V4.1 Flash released?
- DeepSeek V4.1 Flash was released by DeepSeek on September 10, 2026.
- How long a context window does DeepSeek V4.1 Flash have?
- DeepSeek V4.1 Flash supports a context window of up to 1M tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Introducing DeepSeek-V4.1-Flash (DeepSeek, 10 Sep 2026) deepseek.com
- DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression (technical report) arxiv.org
- DeepSeek-V4.1-Flash-1017 model card (Hugging Face, 17 Oct 2026) huggingface.co
- DeepSeek-V4.1-Flash model card (Hugging Face) huggingface.co
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| GLM-5.3 Z.ai (Zhipu AI) | 753B 40B active | 1M | GLM-5.3 License (Conditional) | 520 GB GPU server | |
| Kimi K3.1 Pro Moonshot AI | 2.8T 104B active | 1M | Kimi K3 License (Conditional) | 1.72 TB Multi-node cluster | |
| Reflection Beam Reflection AI | 501B 23.4B active | 1M | Apache 2.0 (Permissive) | 323 GB Big-memory desktop | |
| GigaChat 3.5 Ultra Sber (GigaChat) | 434B – 438B 28B active | 256K | MIT (Permissive) | 299 GB Big-memory desktop |