[UPDATE] DeepSeek V4 Arrives with 1.6 Trillion Parameters and an MIT License

DeepSeek released its V4 model family today, open-sourcing two Mixture-of-Experts variants under the MIT license and launching an API service simultaneously. The release confirms the broad strokes of earlier reporting, including the unprecedented parameter count and the use of Huawei's Ascend chips, while adding one detail that caught many observers off guard: the license is MIT rather than the Apache 2.0 that had been widely expected, making it more permissive for commercial use.

Two models, one release

DeepSeek-V4 comes in two variants. The Pro version has 1.6 trillion total parameters with 49 billion activated at inference time. The Flash version weighs in at 284 billion total parameters, with 13 billion activated. Both support a 1-million-token context window and use a hybrid attention architecture combining what DeepSeek calls Compressed Sparse Attention and Heavily Compressed Attention.

The efficiency gains are significant. According to DeepSeek, V4-Pro requires only 27% of the single-token inference FLOPs that DeepSeek-V3.2 needed at a 1-million-token context length, and only 10% of the KV cache. Running long-context inference is expensive; those reductions matter for anyone deploying the model at scale.

Pricing and access

The API is priced at $0.30 per million tokens. DeepSeek has consistently undercut Western competitors on pricing since its V3 release, and V4 continues that pattern. Model weights are available on Hugging Face.

Performance

On SWE-bench, which measures the ability to resolve real-world software engineering issues, DeepSeek-V4-Pro scores 81%, placing it among the highest-performing models currently available. DeepSeek says the model outperforms Claude Sonnet 4.5 across several coding and instruction-following tasks. Independent verification will follow in the coming days.

The Huawei hardware angle

Earlier coverage noted that V4 was being trained on Huawei Ascend 950 chips rather than Nvidia GPUs, a consequence of US export controls restricting advanced semiconductor sales to China. That reporting holds up: the model was built without Nvidia hardware. For end users, this changes nothing about performance. The broader point is that Chinese AI labs are now demonstrating they can produce frontier-competitive models on domestically produced chips, which is a different kind of data point than anything a benchmark measures.

Full model cards and weights are on Hugging Face. Detailed specs and benchmark analysis at NxCode and OFox AI.