Skip to content
NewsModel release

Moonshot AI ships Kimi K3.1 Pro open weights with a 1 million token context window

· by

The short version

Moonshot AI released open weights for Kimi K3.1 Pro on Hugging Face on November 4, 2026, a 2.8 trillion parameter mixture-of-experts model whose own benchmark table shows gains on hallucination tests alongside strong offensive-security scores.

Moonshot AI ships Kimi K3.1 Pro open weights with a 1 million token context window

Kimi K3.1 Pro's own benchmark table shows a model that hallucinates less and finds software exploits well in the same breath, and Moonshot AI just put both of those capabilities into open weights anyone can download. The release landed on Hugging Face on November 4, 2026.

What

Co-founder and CEO Yang Zhilin's lab had promised the weights when it introduced Kimi K3.1 Pro in October, and November 4 is when that promise was kept, per RuntimeWire. Under the hood sits a mixture-of-experts system: 2.8 trillion total parameters, 104 billion activated per inference, spread across 93 layers and 896 experts, per the model card. The context window runs to 1,048,576 tokens, roughly a million, per TechNode. Training also expanded for coding, 3D creation, and financial research. Long-document and agent tasks got dedicated runs too, per TechNode. On the architecture side, the model card says Moonshot swapped the full-attention layers from the original Kimi K3 for Gated DeepSeek Sparse Attention layers, a change meant to make long-context inference more efficient. A Kimi-specific license governs the release rather than a fully permissive one. Huawei has already announced inference support through vLLM Ascend and SGLang on its Atlas A3 and Ascend 950 systems, per TechNode.

Self-hosted copies carry real offensive-security capability, not just better accuracy

A security team evaluating self-hosted Kimi K3.1 Pro for code review or red-teaming is picking up the same download that hands any other operator competitive exploit-discovery performance once those weights leave Moonshot's platform. The model's own benchmark table backs that up on both sides of the ledger. On Moonshot's internal hallucination test, where lower is better, Kimi K3.1 Pro scores 8.3, ahead of Claude Opus 5.5's 15.0 and GLM-5.3's 22.6, though GPT-6 Astra still leads at 5.5, per the model card. On offensive-security tasks, the same model posts 74.7 on ExploitBench and 81.0 on the internal xdev-bench-hard vulnerability-discovery test, both well ahead of GLM-5.3's 54.4 and 57.1. These are Moonshot's own reported numbers, not an independent audit, but they are the numbers that travel with the weights once anyone self-hosts them, without the usage controls Moonshot applies to its hosted API.

What to watch next

Independent replication of the hallucination and exploit-discovery numbers is the next milestone, since every figure above comes from Moonshot's own model card. Operators should treat those exploit-discovery scores as an unverified benchmark result until that replication lands, and hold off building any production security workflow around them before then.

Sources