Moonshot AI 發佈 Kimi K3.1 Pro，約有 3 trillion parameters，定價約為 GPT-6 Astra 的 1/5，並承諾公開權重

n the early hours of October 24 Beijing time, Moonshot AI released Kimi K3.1 Pro, a post-trained upgrade of its July model Kimi K3 that lands at roughly 3 trillion parameters with a 1-million-token context window — making it, again, the largest open-weight model on Earth by parameter count. The headline is not the size, though. It is the positioning: on Artificial Analysis's intelligence index, K3.1 Pro ranks fourth worldwide and first among all open models, and Moonshot is pricing standard-tier inference at 13 yuan per million input tokens and 68 yuan per million output tokens — about one-fifth the price of GPT-6 Astra — while promising to publish the full model weights by November 4. In financial research benchmarks the company says it beats frontier closed models outright. Combined with fresh OpenRouter data showing Chinese models have topped global weekly token consumption for 25 consecutive weeks and hold the entire top five, this release reads less like a product launch and more like a declaration that the open-source frontier is now a Chinese export.

What actually happened
Kimi K3.1 Pro is the K3.1 series' first release and a heavily post-trained version of July's 2.8-trillion-parameter Kimi K3, with a large increase in reinforcement-learning compute across hundreds of domains. Moonshot's stated focus is exactly where the money is: coding, 3D creation, financial research, and long-horizon agentic tasks — the multi-step, tool-calling, hours-long work that enterprises actually pay for. The company reports that on coding, agent and knowledge benchmarks, K3.1 Pro has reached or approached the level of the strongest closed models, and that in two international financial evaluations it exceeds frontier closed models end-to-end, from document retrieval and cross-document data checking to financial model construction and profitability analysis.

The pricing move is as important as the capability move. At 13 yuan per million input tokens and 68 yuan per million output, K3.1 Pro standard undercuts the previous generation by 35 and 32 percent respectively, and Moonshot publicly frames it against GPT-6 Astra at roughly one-fifth the cost. J.P. Morgan Asset Management's observation — that agent applications inflate token consumption, making total task cost, not sticker price, the real variable in model selection — is precisely the argument open-weight Chinese models are built to win.

The walking humanoid that stole the show
The demo that made the rounds inside the industry is almost poetic. Researchers asked K3.1 Pro to build a simulated human walking system. The model iterated: analyze the experiment, locate the problem, tune the walking controller, retrain. The simulated human went from collapsing within a few steps to alternating strides and walking continuously for 11.2 seconds — and along the way the model identified and corrected 36 errors in the shoulder-muscle force-arm calculations before retraining the controller. It is a small anecdote inside a benchmark sheet, but it is the clearest illustration yet of what "long-horizon agent" means in practice: not answering a question, but running a research loop on its own physics problem for as long as it takes.

The 3D pipeline is similarly agentic: give the model a text description, a sketch, or a reference image, and it infers shape, proportion and spatial relationships, drives professional modeling software, assembles a usable 3D model, and builds it into a full interactive scene — previewable, editable, and downloadable for 3D printing. The coding story follows the same shape: the model wields professional design and development tools, combines web search with 3D generation, and produces complete game and web experiences with noticeably better composition, layout, animation and polish.

The part nobody expected: scrubbing cyber skills on purpose
Buried in the announcement is a governance decision that deserves more attention than it got. Moonshot disclosed that during development it found the model's cyber offense and defense capabilities rising naturally with training scale — no special data required — and that it identified and removed cyber-related data at both the mid-training and reinforcement-learning stages, with such controls to continue at every future stage. The team then published two explicit openness principles: it will not release open weights that push the absolute frontier of cyber offense, and it will gradually raise the level of cyber capability available in open weights. The rationale is worth quoting in spirit: powerful models released openly and runnable on user-controlled infrastructure let far more teams audit the code they maintain, so diffusing the capability beats concentrating it in a few closed systems.

Put that next to this week's PHP2Shell fire — where an OpenAI model found two CVSS 9.2 flaws and a leak turned the patch window into a live-fire exercise — and you can see the industry's new, unspoken bargain forming in real time. One lab strips cyber data out of its open weights and says why; another's model finds the bug of the decade and watches it leak. Both are correct responses to the same discovery: at 3 trillion parameters, cyber skill is no longer a feature you add. It is an emergent property you have to govern.

Why this is exciting
The excitement is structural, not just numeric. For 25 straight weeks, Chinese models have out-consumed American models on OpenRouter — 164 trillion tokens globally in the week of October 12-18, with the entire global top five being Chinese — and K3.1 Pro is the sharpest instrument of that trend yet: open weights, million-token context, frontier-adjacent intelligence, and pricing that forces every closed vendor's CFO to recalculate. Open weights also compound: developers contribute evaluations, tooling and optimizations; enterprises contribute domain data and engineering practice; the ecosystem feeds the next training run. When the largest open model in the world is also the cheapest serious option for long-horizon agent work, the "high-value market" for coding and financial research stops being a closed-lab moat and becomes a contested market — which is exactly where prices fall and capability compounds.

Why it is frustrating — if you sell closed models
For closed-source vendors, the frustration is arithmetic. A rival ships a 3-trillion-parameter open-weight model at one-fifth of GPT-6 Astra's price, beats frontier closed models on two financial evaluations, and promises full weights in eleven days. The Shanghai Securities Press framing is blunt: performance up, cost down, ecosystem expanding — the three reinforce each other, and for closed vendors chasing financing or listings, the question is no longer whether the technology leads, but whether the technology leads to revenue. Enterprises can keep closed models for the hardest work and route high-frequency standardized tasks to cheaper open weights, which means the fight is not just for new customers but for share of the same customer's budget. Every percentage point of that shift is a multiple-point re-rating somewhere in the private markets.
Sources: China News Service (October 24 2026), CGTN, Beijing Daily, CCTV Finance, Shanghai Securities News / China Securities Journal, OpenRouter, threerouter