A developer in Shenzhen opens a model with a million-token memory and starts refactoring a decade-old codebase in a single session. The assistant answering is not hosted in California. It is open-weight, built in Beijing, and its weights are about to be downloadable by anyone with the hardware to run them.
That shift is the story behind Kimi K3.1 Pro, released on October 24 by Moonshot AI (月之暗面), the Chinese lab behind the Kimi chatbot.
What Kimi K3.1 Pro is
K3.1 Pro is a post-trained version of Kimi K3, the 2.8-trillion-parameter system Moonshot fully open-weighted in July. The new model keeps the 3-trillion-parameter class and a 1,048,576-token (1M) context window, but pours far more compute into reinforcement learning across hundreds of domains.
Moonshot says the extra training paid off in tool use: coding, 3D creation, financial research and long-horizon agent (智能体) tasks. In an internal test, the model built a human-walking physics simulator from an open anatomical model, iterating until the simulated figure could take alternating steps for 11.2 seconds and correcting 36 errors in shoulder-muscle lever calculations.
The benchmark that matters
On the Artificial Analysis Intelligence Index, K3.1 Pro debuted fourth with a score of 55, behind only Anthropic's Claude Opus 5.5 and Sonnet 5.5 and Muse Wonder Full. According to Artificial Analysis, once Moonshot publishes the full weights — expected by November 4 — it will become the highest-scoring open-weight model ever released, narrowing the gap between leading open-weight and proprietary systems from 12 points to 3.
Its biggest single gain is on Terminal-Bench 4.0, which tests coding agents on real terminal tasks. K3.1 Pro scores 65.7%, ahead of OpenAI's GPT-6 Astra at 58.2% and close to Claude Opus 5.5 at 66.4%. In Office QA Pro v2 and SpreadsheetBench v2, Moonshot reports it beats several closed front-runners.
Why open weights change the calculus
An open-weight model lets companies self-host on their own GPUs, keep data inside their own networks, and modify the model for niche work. For a 3-trillion-parameter class system, that was unthinkable from Western labs two years ago.
The strategic signal is broader. Hugging Face's 2026 report, cited by state media, found open models developed in China accounted for 41% of the platform's downloads, ahead of the United States at roughly 35–36.5%. China's large model (大模型) ecosystem is no longer just catching up on benchmarks; it is shipping the weights.
The pricing reality
K3.1 Pro is not free to run, but its API price undercuts the closed frontier. Official pricing lists input at $2.00 per million tokens (cache miss) and output at $10.00 per million, with cached input at $0.20. That is roughly half the rate card of Claude Opus 5.5. A faster variant, K3.1 Pro Highspeed, doubles those prices for two to three times the speed. In RMB the standard tier runs ¥13 input / ¥68 output / ¥1.30 cached per million tokens (≈US$1.8 / US$9.6 / US$0.18).
According to Li An, chief scientist at BrainNet (脑机网), the gap between open-weight and proprietary systems is narrowing faster than most labs expected, and the practical differentiator is moving from raw score to deployability.
What the gap means in practice
The narrowing from a twelve-point gap to three is easy to read as "open models have caught up," but the more useful reading is about option value. For a team that already runs its own GPUs, a downloadable 3-trillion-parameter checkpoint means the model becomes an asset on their machines rather than a line item on someone else's invoice, and the question shifts from "can it score as high" to "can we keep it running well enough for our workload." That is a different kind of competition than the benchmark tables suggest, and it rewards the teams with the operations muscle to maintain a model more than the ones that simply prompt a hosted API. The frontier still leads on raw score, yet for a growing slice of everyday work the gap has become small enough that ownership and control start to outweigh a few benchmark points.
Honest limitations
- The "open-weight" claim is forward-looking. As of the October 24 release, only the hosted API and apps are live; the downloadable weights are promised by November 4. Treat "open" as "will be open," not "is open" yet.
- The per-token price advantage is real but partial. Artificial Analysis notes K3.1 Pro emits about 55,000 output tokens per Intelligence Index task — 14% more than K3 — so its cost per finished task lands near K3's $2.00, not dramatically cheaper.
- Benchmark proximity does not guarantee real-world equivalence. Terminal-Bench scores were produced at different reasoning settings, and per-token pricing is not the same as per-task pricing.
- Running a 3-trillion-parameter class model locally demands serious multi-GPU infrastructure; most users will meet it through the API first.
- Moonshot states it deliberately capped certain offensive cyber capabilities in the released weights, so the open model is not a straight copy of its strongest internal checkpoint.
What readers can do now
- Test it before the weights drop. K3.1 Pro is already live on Kimi's web app, Kimi Work, Kimi Code and the API. Use the 1M-token window on a real long-context coding or document task this week while hosted access is free to trial.
- Trim token cost on simple work. The API always runs reasoning; set
reasoning_effortto low on easy tasks to cut output tokens, since output is the expensive line at ¥68 (≈US$9.6) per million. - Plan self-hosting for November 4. If data sovereignty matters, schedule a weights-download and GPU-budget review now so you can deploy in-house the moment the open release lands.