Kimi K3.1 Pro Weights Released: 2.8T Parameters Explained

By GameDrop Staff · Nov 05, 2026 · AI · 2 min read

Image: Moonshot AI

Moonshot AI has put the weights of Kimi K3.1 Pro on Hugging Face, so developers can now run its flagship multimodal model themselves. TechNode reported the release on November 5, 2026.

Quick summary

Architecture: mixture-of-experts with 2.8 trillion total parameters and about 104 billion active per inference.
Context window of around one million tokens.
Built on Kimi K3 with more reinforcement learning for coding, 3D creation, financial research and long-running agent tasks.
Weights were released after the model was already available in Kimi's apps and API.
Huawei announced inference support through vLLM Ascend and SGLang on its Atlas A3 and Ascend 950 systems.

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is Moonshot AI's flagship multimodal model. The "active parameters" figure matters: only about 104 billion of the 2.8 trillion parameters are used for each token, which is how huge mixture-of-experts models keep inference costs manageable.

What changed from Kimi K3?

According to TechNode, K3.1 Pro expands reinforcement learning for four areas: coding, 3D creation, financial research and long-running agent tasks. No benchmark scores were included in the report we checked, so we are not quoting any.

Can you run Kimi K3.1 Pro at home?

Realistically, no. A 2.8-trillion-parameter model needs a multi-GPU server, even though only part of it is active per token. Most people will use it through the Kimi app or API. Open weights mainly matter for companies and researchers who want to self-host or fine-tune. Huawei's inference support shows it is aimed at server hardware, including non-Nvidia chips.

How does it fit the wider AI race?

It lands in a busy week for open-weight models. Tencent launched Hy4, a 770-billion-parameter model with vision and computer use, on November 4 (see our Tencent Hy4 explainer), and Baidu open-sourced ERNIE 5.5 Flash Base, a 280B model with 14B active, on November 5. Meituan's LongCat-2.5 arrived earlier in the week, covered in our LongCat-2.5 guide.

Model	Total params	Active params	Notes
Kimi K3.1 Pro	2.8T	~104B	~1M context, weights on Hugging Face
Tencent Hy4	770B	Not stated here	Vision and computer use, weights promised
Baidu ERNIE 5.5 Flash Base	280B	14B	Base model, 1M context in fine-tuning

Verdict

For most readers this is a "watch" story, not a download-now one. The practical takeaway is that frontier-scale open weights keep coming from China, and independent benchmarks will show whether K3.1 Pro earns its size.