Kimi K3.1 Pro opens a frontier model’s weights and cuts its cyber side
Moonshot AI releases the weights of Kimi K3.1 Pro, a 2.8-trillion-parameter MoE with 104 billion active parameters and a 1-million-token context, which beats GPT-6 Astra on Terminal-Bench 4.0 at 65.7%. Before adopting it, read the license — the Kimi name is mandatory for large products — and know that cyber capabilities were stripped from the training data.
November 4, 2026. Moonshot AI publishes the weights of Kimi K3.1 Pro, the first model in its K3.1 series and its most capable open-weight model in the 3-trillion-parameter class, with a 1-million-token context window. November 5, 2026. A Highspeed variant, two to three times faster, is announced. Why it matters: an open-weight model now sits between GPT-6 Astra and Claude Opus 5.5 on coding-agent benchmarks — but “open weights” does not mean “open source,” and the vendor has explicitly removed cyber capabilities from its training data.
A 2.8-trillion MoE built for agents
Kimi K3.1 Pro is a MoE (mixture of experts) with 2.8 trillion total parameters, of which 104 billion are active per inference, distributed across roughly 1.56 TB in MXFP4. The one-million-token context and native image input put it in the same league as proprietary frontier models — but with downloadable weights.
The positioning shows in the numbers. On Terminal-Bench 4.0, which evaluates coding agents on realistic terminal tasks under strict per-task tests, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%). On Kimi Code Bench 3.0 Long-horizon, its internal multi-step agent benchmark, it reaches 61.2%, against 55.95% for its predecessor K3. In 3D, on BenchCAD, it hits 92.7% with tools, close to Opus 5.5 (96.2%) and GPT-6 Astra (95.9%).
The gain comes from heavy post-training. Moonshot describes optimizing “every layer of the stack” — the RL framework, long-context inference, sandbox infrastructure, rollout scheduling — so that each increment of RL compute yields predictable returns across “hundreds of downstream domains.” The practical translation: the model was trained to execute long agentic tasks, not merely answer questions.
A $4-to-$20-per-million-token price tag
Price is the other argument. Moonshot charges $4 per million input tokens and $20 per million output tokens, a fraction of proprietary frontier pricing. For a team that already runs GPUs and wants to avoid API lock-in, it is an economically legible option — provided you accept the hosting cost of a 1.56 TB model in reduced precision.
The Highspeed variant, two to three times faster, targets interactions where latency beats depth. The model is available on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, in addition to the downloadable weights on Hugging Face.
# Serving K3.1 Pro locally requires a multi-GPU fleet that can hold
# ~1.56 TB of MXFP4 weights; a single 80 GB GPU is not enough.
# Example check (adapt to your inference engine):
nvidia-smi --query-gpu=name,memory.total --format=csv “Open weights,” not open source: read the license
This is where integrators should slow down. The weights are open, but not in the open-source sense. Two clauses change the picture for a commercial product: large products must display the Kimi name, and large-scale API resellers need a separate agreement. It is a hybrid license model — more open than a closed API, more constrained than a permissive license — in line with what other open-model vendors now practice.
The operational consequence is direct: before building a product on K3.1 Pro, the legal team must read the license, not just the benchmark. A model you can download but not freely redistribute under your own brand is not neutral in a product plan.
Cyber removed from the corpus, on purpose
The most unusual passage in the announcement concerns cybersecurity. Moonshot explains that it watched vulnerability-research capability improve steadily as post-training scaled, and traced the trend to its source: cyber-related content was embedded in the mid-training corpus, and cyber-related tasks appeared in the RL environment mix. For the open-weight release, the vendor identified and removed that data from both stages, while keeping the general coding tasks the model needs.
The decision is an explicit trade-off between capability and control. By stripping cyber from the corpus, Moonshot lowers the misuse risk of a freely downloadable model, at the cost of a capability a closed frontier model keeps. It is a measured answer to the question every open-weight vendor now has to settle: what to do with dual-use capability when you open the weights. The hallucination numbers point the same disciplined direction: on Kimi Hallu Bench Hard, the claim-level error rate falls from 23.4% for K3 to 8.3% for K3.1 Pro.
Verdict
If you already run a GPU fleet that can hold 1.56 TB of MXFP4 weights and you want an alternative to frontier APIs for coding agents, K3.1 Pro is a serious option: it beats GPT-6 Astra on Terminal-Bench 4.0 and costs $4 to $20 per million tokens through the API. If you are building a commercial product, get the license reviewed before integrating — the mandatory Kimi-name display for large products is not trivial, and API resellers have a separate regime. If you were counting on this model for offensive vulnerability research, look elsewhere: cyber capabilities were deliberately removed from training. Moonshot’s announcement is not just another benchmark — it is a new balance between open weights, price, and capability control, and it is that balance you should evaluate, not the spec sheet alone.