DeepSeek V4

DeepSeek's MIT-licensed 1M-context mixture-of-experts pair: a 1.6T Pro and a 284B Flash, built for long-context agents.

In short. DeepSeek V4 is a pair of very large free AI models, Pro and Flash, that can read about a million tokens at once without the usual memory bill. They are aimed at coding and agent work, and the MIT license lets you use them for anything. They are server-scale models, not something for a laptop.

Summary

DeepSeek announced the V4 Preview on 24 April 2026 with two mixture-of-experts models: V4-Pro with 1.6 trillion parameters (49 billion active per token) and V4-Flash with 284 billion (13 billion active). Both have a one-million-token context window, which DeepSeek made the default across its own services, and both use the MIT license for code and weights. The previews were later replaced by official checkpoints, V4-Flash-0731 on 31 July and V4-Pro-0813 on 13 August 2026. DeepSeek says they supersede the previews with much stronger agentic abilities, and each includes a built-in speculative-decoding module called DSpark, so no separate draft model is needed.

The big change from DeepSeek V3.2 is how attention copes with long inputs. V4 combines Compressed Sparse Attention and Heavily Compressed Attention, and the technical report says that at one million tokens V4-Pro needs about 27% of V3.2's per-token compute and 10% of its KV cache. The models also add Manifold-Constrained Hyper-Connections, which strengthen the residual connections, and were trained with the Muon optimizer on more than 32 trillion tokens. V4-Flash has 43 layers with 256 routed experts plus one shared expert, and V4-Pro has 61 layers with 384 routed experts; both route each token to 6 experts.

The instruct checkpoints store the expert weights in FP4 and most other weights in FP8, which comes to about 160 GB for the V4-Flash preview and 865 GB for the V4-Pro preview. The official 0731 and 0813 checkpoints are larger, about 167 GB and 893 GB, because they include the DSpark module. Three reasoning modes are available, Non-think, Think High and Think Max, and DeepSeek recommends a context window of at least 384K tokens for Think Max. There is no Jinja chat template; DeepSeek ships Python encoding scripts instead. In DeepSeek's own tables the preview V4-Pro in Max mode scores 80.6 on SWE-bench Verified and 67.9 on Terminal-Bench 2.0, and the later releases report Terminal-Bench 2.1 scores of 82.7 for Flash-0731 and 87.9 for Pro-0813. All of these figures are self-reported.

Good for

  • Long-context coding and agent sessions that need hundreds of thousands of tokens in one conversation
  • Teams with multi-GPU servers that want a very large open model under the plain MIT license
  • Using V4-Flash as the cheaper step down, with 13B active parameters per token
  • Products that cannot accept revenue limits or usage restrictions

Watch out for

  • Even V4-Flash-0731 is about 167 GB, and V4-Pro-0813 is about 893 GB, so both need a multi-GPU server or a very large unified-memory machine
  • The main checkpoints read and write text only; image input needs DeepSeek V4.1 Flash or the experimental V4-Flash-Vision-Exp
  • The previews and the official 0731 and 0813 releases are different checkpoints with different scores, so check which one a benchmark or provider means
  • Benchmark figures come from DeepSeek's own tables, and the agent scores use its own harness

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the V4-Flash (preview) 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-Flash-0731 FP4 284B 13B active 653 GB 327 GB 184 GB Big-memory desktop
V4-Flash (preview) FP4 284B 13B active 653 GB 327 GB 176 GB Big-memory desktop
V4-Pro-0813 FP4 1.6T 49B active 3.68 TB 1.84 TB 982 GB GPU server
V4-Pro (preview) FP4 1.6T 49B active 3.68 TB 1.84 TB 952 GB GPU server

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

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.

More about MIT

See also

Frequently asked questions

What is DeepSeek V4?
DeepSeek V4 is a pair of very large free AI models, Pro and Flash, that can read about a million tokens at once without the usual memory bill. They are aimed at coding and agent work, and the MIT license lets you use them for anything. They are server-scale models, not something for a laptop.
Can I use DeepSeek V4 commercially?
Yes. DeepSeek V4 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 open source?
DeepSeek V4 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 on my own computer?
Only on a machine with a lot of memory. The smallest variant, V4-Flash (preview), needs about 176 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was DeepSeek V4 released?
DeepSeek V4 was released by DeepSeek on April 24, 2026.
How long a context window does DeepSeek V4 have?
DeepSeek V4 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.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
DeepSeek V4.1 Flash DeepSeek 552B 16B active 1M MIT (Permissive) 561 GB GPU server
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
Qwen3.8-Max Alibaba Qwen 2.45T 95B active 1M Qwen3.8-Max License (Conditional) 1.69 TB Multi-node cluster