Moonshot publishes Kimi K3.1 Pro’s weights: a 2.8T open model that nearly matches Claude Opus 5.5 on coding

Moonshot AI has handed the keys to its biggest model to everyone. On November 4, the Chinese lab published the full weights of Kimi K3.1 Pro on Hugging Face, twelve days after the model went live — on the exact date it had promised. At 2.8 trillion parameters, it is the largest open-weight model ever released, and Moonshot's own numbers put it within a point of Claude Opus 5.5 on the agentic coding test developers actually care about.

The release matters less for what you can run at home — almost nobody can host a 1.56-terabyte download — than for what it does to the market. A model that close to the frontier is now downloadable, inspectable, and servable by any provider with enough hardware, under terms you can read in advance. On Artificial Analysis's Intelligence Index, K3.1 Pro scores 55, three points behind Claude Opus 5.5 — and The Decoder estimates that cuts the open-versus-proprietary gap on that index from twelve points to three.

What actually shipped#

The repository, moonshotai/Kimi-K3.1-Pro, holds 96 weight files totalling roughly 1.56 TB. K3.1 Pro is a mixture-of-experts model: 2.8 trillion parameters in total, with 104 billion active for any given token, spread across 896 experts, 16 of which are consulted per token. The context window is 1,048,576 tokens, and it reads text and images natively through a 401-million-parameter vision encoder. The weights ship in MXFP4, a 4-bit format the model was trained to tolerate from the start, which is why the download is 1.56 TB instead of the roughly 5.6 TB a 16-bit copy would need.

It is not a new base model. Moonshot describes K3.1 Pro as Kimi K3 — launched in July — post-trained with substantially more reinforcement-learning compute, tuned jointly across hundreds of downstream domains. It has been live since October 23 in Kimi’s apps and API at $2 per million input tokens and $10 per million output tokens; the weights were the missing piece.

Running it is a data-center job. Moonshot recommends the vLLM, SGLang or TokenSpeed inference engines, and the model is also available through the Kimi API and Hugging Face inference providers. Few organisations will self-host; the practical audience is inference providers and labs.

The numbers, and how much to trust them#

Moonshot's headline number is Terminal-Bench 4.0, a test of real agentic coding work: K3.1 Pro scores 65.7, against 66.4 for Claude Opus 5.5 and 58.2 for GPT-6 Astra. It also leads Opus 5.5 on SWE Marathon (51.7 to 46.3) and on Agents' Last Exam (43.2 to 42.6), and trails narrowly on the Artificial Analysis knowledge-work Elo, 1846 to 1866, and on OSWorld computer use, 79.6 to 81.8.

But the table is not a clean sweep. On Terminal-Bench Science it scores 38.1 against Opus 5.5’s 63.3; on HLE-Diamond, a hard knowledge test without tools, it manages 39.4 against GPT-6 Astra’s 65.8. This is a coding-and-agents specialist that is merely good elsewhere.

And every number in the table is Moonshot's own. The model was evaluated with the Claude Code harness, competitor scores were compiled rather than re-run, and several of Moonshot’s benchmarks are internal. The one outside number so far comes from Artificial Analysis: 55 on its Intelligence Index, fourth behind Opus 5.5 at 58, and the highest an open-weight model has reached there. Independent runs on the released weights will take time; until then, treat the vendor table as a signal, not a verdict.

The license has teeth for big operators#

The weights ship under the Kimi K3.1-Pro License, which reads like an MIT license with two conditions aimed squarely at very large operators. A company that runs a model-as-a-service business and earns more than $20 million over any twelve months must sign a separate agreement with Moonshot before commercial use. And a product with more than 100 million monthly users, or more than $20 million in monthly revenue, must display "Kimi K3.1-Pro" in its interface.

For almost everyone else — researchers, startups, internal enterprise use — the license is effectively permissive: download, fine-tune, and serve freely. The conditions bite only at the scale where a provider could build a business reselling Moonshot’s model without Moonshot.

The cybersecurity dial#

The most unusual part of the release is what Moonshot deliberately left out. Moonshot says it traced part of that climb to cyber-related content in the training corpus and reinforcement-learning environments, and removed that data at both stages for the open-weight release. On its internal exploit benchmark, the score fell from 90.5 for the unreduced checkpoint to 81.0 for the model as released — and on the public ExploitBench it scores 74.7, against 100 for GPT-6 Astra.

Moonshot states two principles behind the decision: not to release open weights that extend the absolute frontier of cyber capability, and to progressively raise the capability available in open weights over time. It is a rare public admission that open-weight labs now set a deliberate dial on dual-use capability. An 81.0 on exploit tasks is still a capable model; teams giving coding agents shell access should sandbox and gate them regardless of which model is underneath.

Why it matters#

Three things change with this release. First, choice: a near-frontier agentic model can now be served by any provider with the hardware, in any region, on published terms. That puts downward pressure on API prices and gives buyers a fallback if a single vendor changes its terms. Second, verification: with the weights public, independent labs can finally test Moonshot’s claims on the actual artifact instead of taking the vendor’s table on trust. Third, the frontier itself: the open-weight field has never been this close to the closed labs on agentic coding.

What to watch: independent evaluations run on the released weights; which inference providers pick it up and at what price; and whether the next open releases keep closing the gap. The weights are out. The verdict is now everyone’s to check.

Sources: Moonshot AI, Kimi K3.1 Pro model card on Hugging Face; ExplainX; Fulcrum Cortex; and CellCog. Images are AI-generated editorial illustrations.