Today, November 4, 2026, Moonshot AI is making good on its promise: the full weights of Kimi K3.1 Pro — a 2.8 trillion parameter mixture-of-experts model — are being released publicly on Hugging Face. The model launched via API on October 23, with the open weight release scheduled for November 4. It is one of the largest open-weight models ever released.
The Architecture
Kimi K3.1 Pro uses a mixture-of-experts design: 2.8 trillion total parameters with 104 billion active per token. That means inference is far cheaper than a dense 2.8T model — you get frontier-class reasoning while only activating a fraction of the network on each forward pass. The model also supports a 1-million-token context window, making it competitive with Gemini and Claude on long-document tasks. DataCamp's breakdown notes it as the first open-weight model to combine this context length with a MoE architecture at this parameter scale.
Benchmarks (Vendor-Reported)
Moonshot's own numbers show the model landing just below Claude Opus 5.5 on coding and reasoning tasks. On Terminal-Bench 4.0, K3.1 Pro scores 65.7% against Opus 5.5's 66.4%. On MMMU-Pro (multimodal reasoning), it achieves 84.1% versus 87.7% for Opus. On ExploitBench (offensive security), Moonshot says it deliberately kept the score lower — at 74.7 — by removing cyber-related training data before the public release, a measure intended to reduce dual-use risk.
These figures have not been independently verified. The community is running evals now that the weights are public, and those numbers may diverge significantly from Moonshot's launch table.
Controversy: The Paid-Promotion Scandal
The launch has not been clean. A DEV Community post detailed evidence that several influential AI reviewers published glowing write-ups without disclosing paid arrangements with Moonshot. The pattern — coordinated posts going live within hours of the October 23 launch, nearly identical phrasing, benchmark tables sourced directly from Moonshot — was spotted by multiple community members cross-referencing timestamps and language. Moonshot has not publicly responded.
Licensing: Still Unknown
The predecessor Kimi K3 shipped under a custom, non-OSI-approved license. Moonshot has not announced the K3.1 Pro terms ahead of the weight release. Whether commercial use is permitted — a question that matters a great deal to the many teams waiting to integrate the model — remains open as of today.
Bottom Line
If the benchmarks hold under independent scrutiny, Kimi K3.1 Pro slots in just below the leading proprietary models at a fraction of the inference cost per token, and with weights you can run yourself. That's a significant outcome for the open-source AI ecosystem. But the missing license terms and the benchmark promotion controversy mean the community should approach the initial numbers with appropriate skepticism — and check Moonshot's Hugging Face page for the official model card before building anything on top of it.