HYPE OR FACT? · CHECKED 4 NOVEMBER 2026
Did Moonshot really release the Kimi K3.1 Pro weights?
OUR VERDICT
FactOur original verdict: “Yes. The weights are public on Hugging Face (not gated) under a modified-MIT "Kimi K3.1 Pro License". The model has 2.8T total and 104B active parameters, with a 1M-token context. Moonshot's own benchmarks put it close to top closed models on agentic and coding tasks, and Artificial Analysis's independent runs broadly agree on the agentic numbers. It is still behind on hard reasoning and cyber tests. We haven't tested it hands-on.”
THE CLAIM
“Moonshot AI says it is "releasing the model weights of Kimi K3.1 Pro" and will "keep advancing open intelligence".”
What’s actually true
- Weights are live and ungated at huggingface.co/moonshotai/Kimi-K3.1-Pro: about 2.78T parameters in safetensors, stored as MXFP4 weights (quantization-aware training).
- Model card specs (Moonshot's figures): mixture-of-experts with 2.8T total and 104B active parameters, a 1,048,576-token context, and text plus image input via a 401M-parameter MoonViT-V2 encoder.
- The licence is MIT-style with conditions. A company running Model-as-a-Service with more than $20M revenue over 12 months needs a separate deal with Moonshot. Products with more than 100M monthly users or $20M monthly revenue must display "Kimi K3.1 Pro". Internal use is exempt from both conditions.
- API (Moonshot's figures, as reported by Artificial Analysis): $2 input, $10 output and $0.20 cache hit per 1M tokens, 33% below Kimi K3. Batch jobs cost 60% of the standard price.
- Benchmarks are Moonshot's own. Examples: Terminal-Bench 4.0 65.7% vs Claude Opus 5.5 66.4%, and HLE-Full with tools 70.0. Several are internal benchmarks (Kimi Code Bench, Kimi Hallu Bench, Business Operation, DECK-Arena) that nobody outside can check.
- Independent: Artificial Analysis measured 67% on Terminal-Bench 4.0 and 73% on AutomationBench-AA. Its launch write-up ranked the model 4th on its Intelligence Index and called it the top open-weights model. Its current model page, on index v5.0.1, shows 46 (#22 of 62) and describes the model as slow and verbose.
- Weak spots on Moonshot's own table: it trails the top closed models on HLE-Diamond (39.4 vs 60-66), Terminal-Bench-Science (38.1 vs 63-68), ProgramBench and the cyber tests. LMArena has no Text score yet.
Sources
Cost
Weights are free to download. The API costs $2 in and $10 out per 1M tokens (Moonshot pricing via Artificial Analysis), and it is also in the Kimi app plans.
USE IT IF
Teams that want a near-frontier agentic and coding model they can self-host, or a cheaper API than the top closed models.
SKIP IT IF
Anyone without multi-GPU datacentre hardware who wants to run it locally (it is 2.8T parameters). Large Model-as-a-Service providers who haven't signed a deal with Moonshot.
How we knowChecked on 4 November 2026. Verdicts: Fact = the claim holds up; Partly fact = real, but with caveats or overstated parts; Hype = the headline claim isn't supported. Products change quickly, so check the maker's site before you buy. Our methods & sources ↗