we_are_coded.by CODE · The world, decoded
БГ

Moonshot AI · Open models

Moonshot publishes Kimi K3.1 Pro weights: the same 2.8 trillion parameters, new attention for long context

According to Moonshot AI, Kimi K3.1 Pro is an open-weight model post-trained from K3: 2.8 trillion parameters, 104 billion active, a context of 1 million tokens. The results table is Moonshot's own, and the terms of the kimi-k3.1-pro license are not in the text we read.

Source
Moonshot AI
Published
The event
Checked on
Responsible editor
Cvetelin Ivanov
How we work
MethodCorrections

The size has not moved. Kimi K3.1 Pro has 2.8 trillion parameters, 104 billion of them active - same as K3. What moved is the attention. According to Moonshot, K3.1 Pro swaps the full-attention Gated MLA layers for Gated DeepSeek Sparse Attention (Gated DSA) layers, for more efficient work with long context. And the context is 1 million tokens.

Fact

The facts: on 6 November 2026 Moonshot AI published the weights of Kimi K3.1 Pro on Hugging Face (the date comes from the repository's machine timestamp; the page shows no visible announcement date). According to Moonshot it is an open-weight, natively multimodal, agentic model, its most capable 3T-class model to date. It is post-trained from Kimi K3 and is the first model in the K3.1 series. The architecture is a mixture of experts: 2.8 trillion parameters in total, 104 billion active, 896 experts, of which 16 are selected per token. Quantization is MXFP4 for weights and MXFP8 for activations (quantization-aware training), and the input is text and images. The context is 1 million tokens. The API is reached through platform.kimi.ai with the model kimi-k3.1-pro. The page lists a license named kimi-k3.1-pro; its terms are not in the text we read.

The table is Moonshot's. Results are at max reasoning effort, and some of the tests are internal. By that table Kimi has the highest number in its row on some of the tests, among them HLE-Full with tools, 70.0, and SWE Marathon v1.1, 51.7. On Terminal-Bench 4.0 it scores 65.7, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. On ExploitGym, the cyber test, Kimi has 20.7 and GPT-6 Astra has 42.4. The picture is uneven, which is normal for a table that picks its own rivals.

Same size, different attention.

If you plan to download it, open the kimi-k3.1-pro license first - its terms are not in the text we read. We have not seen an independent check of the numbers yet. Until then, the numbers are Moonshot's.

The visual is generated code-art, no third-party images.