Kimi K3.1 Pro becomes downloadable with open weights

Moonshot AI has released the weights of Kimi K3.1 Pro on Hugging Face. The cutting-edge multimodal model can now be deployed independently, with a context window of about one million tokens.

Moonshot AI has released the weights of Kimi K3.1 Pro on Hugging Face, making this flagship multimodal model available for independent deployment. Behind this announcement is a mixture-of-experts architecture totaling 2.8 trillion parameters, with about 104 billion activated per inference, and a context window of about one million tokens. This move is significant for teams wanting to move beyond a single API framework while retaining access to a model designed for coding, financial research, and long-duration agent tasks.

A more freely deployable model, but not lighter

The weights of Kimi K3.1 Pro were previously accessible via the Kimi apps and its API. Publishing on Hugging Face changes the game for internal uses, as it allows autonomous exploitation of the model, subject to usual hardware constraints. In short, access broadens, but the model does not fit on a small desktop machine: with a mixture-of-experts architecture, only part of the parameters is activated per request, which reduces inference cost without eliminating the model's enormity.

Moonshot AI presents K3.1 Pro as an evolution of Kimi K3, enhanced by extended reinforcement learning for coding, 3D creation, financial research, and agents capable of chaining long tasks. This positioning targets professional uses where context length and reasoning stability matter as much as response speed.

Huawei connects to inference

Huawei also announced support for inference via vLLM Ascend and SGLang on its Atlas A3 and Ascend 950 systems. This compatibility sends a useful signal for the Chinese enterprise AI market: open or semi-open models no longer live only in research catalogs, they also enter the hardware and software stacks of major providers. In this context, Kimi K3.1 Pro gains an additional distribution channel beyond the Moonshot ecosystem.

The usual question remains, less glamorous than the numbers but decisive for buyers: actual availability, operating cost, and maturity of integrations [to be verified]. The model promises a lot on paper. In practice, it is often the electricity bill, inference times, and orchestration simplicity that decide if AI moves from test to production.

What this release changes for AI teams

For companies and labs wanting to keep control over their deployments, opening the weights is a strong signal. It facilitates local testing, fine-tuning behaviors, and compliance in sensitive environments. It also pushes competition to take a closer look at a Chinese model capable of playing in the big leagues across multiple tasks simultaneously.

The K3.1 Pro case illustrates a broader shift: frontier models are no longer just consumed via an interface or remote API. They also become building blocks that are installed, optimized, and supervised, with all that implies in terms of technical governance and infrastructure cost.

Key points

  • Moonshot AI publishes the weights of Kimi K3.1 Pro on Hugging Face.
  • 2.8 trillion parameters in total, 104 billion activated per request.
  • Context window of about one million tokens.
  • Huawei supports inference via vLLM Ascend and SGLang.
  • K3.1 Pro extends Kimi K3 for code, 3D, and finance.
  • API availability preceded this weight release.

By the numbers

  • 2.8 trillion parameters — total model size, according to Moonshot AI.
  • 104 billion parameters — parameters activated per inference, according to Moonshot AI.
  • 1 million tokens — announced context, according to Moonshot AI.
  • 2 systems — Atlas A3 and Ascend 950, on Huawei's side.

To read