Model releases · 11 August 2026
A sparse model released under one of the most permissive licences in the field, shipped already quantised to FP8, and built around a paper whose title is about the cost of long context rather than raw capability.
| Released by | DeepSeek |
|---|---|
| Published | 31 July 2026 (0731 checkpoint) |
| Total parameters | ~167B, inferred from the FP8 weight size |
| Activated parameters | Not stated in the model card |
| Architecture | Mixture-of-Experts (DeepseekV4ForCausalLM) |
| Layers / hidden size | 43 / 4096 |
| Experts | 256 routed, top-6 per token, plus 1 shared |
| Context length | 1,048,576 tokens (1M) |
| Vocabulary | 129,280 tokens |
| Licence | MIT |
| Weight format | safetensors, FP8 quantised — 166.9 GB across 48 shards |
| Hosted at | deepseek-ai/DeepSeek-V4-Flash-0731 on Hugging Face |
DeepSeek V4 Flash is the smaller, faster member of the V4 line. DeepSeek's own claim is that this checkpoint outperforms DeepSeek-V4-Pro (Preview) on their published benchmarks despite a far smaller activated parameter count, and is broadly competitive with the strongest proprietary models.
The accompanying paper is titled DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence — the framing is about making very long context affordable rather than about scaling capability.
A detail worth noticing: the weights are published in FP8 rather than bfloat16. The configuration declares bfloat16 as the compute dtype with an FP8 quantisation config on top. That is why 166.9 GB corresponds to roughly 167 billion parameters rather than half that — and it means there is no separate "download the full precision version" step.
256 routed experts with the top 6 selected per token, plus one shared expert always active, across 43 layers at hidden size 4096. The MoE intermediate size is 2048. The vocabulary is 129,280 tokens, notably smaller than the 200k-plus vocabularies in several competing releases.
Over a million downloads and 3,100 likes make this one of the most-adopted open releases of the period, and the MIT licence is a large part of that: no acceptable-use appendix, no regional restriction, no commercial carve-out.
DeepSeek also publishes a related DeepSeek-V4-Flash-DSpark variant, alongside a set of small speculative-decoding drafters trained against other base models.
| Runtime | Supported | Notes |
|---|---|---|
| transformers | Yes | Standard configuration published by the lab. |
| llama.cpp | Partial | Support for the DSpark and multi-token-prediction paths landed in llama.cpp after release. Community GGUF conversions exist, produced by third parties rather than DeepSeek. |
| MLX | Partial | Community conversions only; nothing published by the lab. |
| Official quantised build | Yes | The published weights are already FP8. |
MIT, with no additional acceptable-use policy. That is among the most permissive terms applied to a model of this size.
166.9 GB across 48 safetensors shards. The weights are published in FP8 rather than bfloat16, so that size corresponds to roughly 167 billion parameters.
1,048,576 tokens — one million. The accompanying paper is specifically about making context at that length efficient.
256 routed experts with the top 6 activated per token, plus one shared expert that is always active.
No. The GGUF builds in circulation were produced by third parties. The lab publishes safetensors in FP8.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026