Moonshot AI has released the weights of Kimi K3.1 Pro on Hugging Face, keeping the promise it made at the October 24 launch, ITHome reported on November 5. The model has 2.8 trillion parameters in total, activates about 104 billion per inference pass, and supports a context window of roughly 1 million tokens.
The size of the weight files is just as striking: the BF16 version is about 5.6TB, and the INT4-quantized version comes to about 1.4TB.
The license is the detail worth waiting on. According to public information, the previous generation, K3, used a modified MIT license. We have not yet confirmed the terms for K3.1 Pro on any official page, so readers planning commercial use should go by the LICENSE file in the repository.
A leaderboard scored on real sessions
The result cited alongside the weights comes from Arena’s Agent Arena leaderboard. K3.1 Pro ranks first among open-weight models and 13th among all models evaluated, and it is the top open model in all three categories: work, coding, and chat.
This leaderboard does not rely on pairwise blind voting. It extracts signals from real agent sessions, such as task confirmations, praise and complaints, user corrections, and shell errors, then estimates a net improvement rate using a randomized inference method called “causal tracing.” This evaluation covered more than 17,500 long-horizon task sessions.
| Metric | K3 | K3.1 Pro |
|---|---|---|
| Net improvement rate | +0.9% | +5.1% |
| Bash reliability | -4.5% | +7.9% |
| Praise vs. complaints | +1.3% | +7.7% |
Bash reliability swung from negative to positive and also took first place among open models. That matters most to people running long tasks: an agent that keeps failing on commands is hard to leave unattended. The task-confirmation success rate of +3.9% ranks fourth among open models, so the model does not lead on every measure.
Against two closed flagships
Moonshot AI’s published comparison table covers dozens of benchmarks, including internal tests. The table below keeps only the public ones.
| Benchmark | K3.1 Pro | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 65.7 | 58.2 | 66.4 |
| SWE Marathon v1.1 | 51.7 | 42.5 | 46.3 |
| Frontier Finance | 58.6 | 45.0 | 56.7 |
| HLE-Full (with tools) | 70.0 | 57.2 | 67.7 |
| Program Bench | 31.5 | 50.0 | 65.0 |
| Terminal-Bench-Science 0.1 | 38.1 | 68.1 | 63.3 |
| HLE-Diamond | 39.4 | 65.8 | 60.9 |
| ExploitBench | 74.7 | 100 | 91.0 |
The tasks where K3.1 Pro leads are mostly tool-using, long-workflow ones. Where it trails are scientific tasks, hard reasoning, and offensive and defensive security. On Terminal-Bench 4.0 it sits just 0.7 points behind Opus 5.5, but on that benchmark’s science branch the gap widens to more than 25 points. The table was published by the vendor, and some of the scores have not been reproduced by third parties.
The security row deserves its own look. Once an open-weight model is downloaded, it cannot be recalled. Scores of 74.7 on ExploitBench and 20.7 on ExploitGym still sit some distance from the closed flagships, and they offer one reference point for the debate over open releases and security.
Downloadable is far from runnable
ITHome specifically cautioned in its report that there is a big gap between “downloadable” and “runnable locally,” and that ordinary consumer GPUs can hardly deploy this model.
A rough calculation: holding the 1.4TB INT4 version entirely in GPU memory takes about 18 data-center cards at 80GB each, before counting the context cache. The MoE architecture activates only about 104 billion parameters at a time, and some full-attention layers have been swapped for gated sparse attention to cut the cost of long contexts, but the weights still have to live somewhere.
For most teams in China, then, the direct benefit of this release falls to a handful of cloud vendors and inference providers, which can deploy the model themselves and set their own prices. What ordinary developers are more likely to notice is more places to call it from. Whether prices come down as a result will have to be compared once third-party hosted versions go live.
Sources: ITHome, CocoLoop, Arena Agent Arena leaderboard; benchmark scores come from the comparison table published by Moonshot AI, and the memory requirements are our own estimate.