Kimi K3.1 Pro Stuns With A 1M-Token Open-Weight Flagship

Moonshot AI has released Kimi K3.1 Pro, the first model of its new K3.1 series and the strongest open-weight model the company has ever shipped: a 3T-class system with a one-million-token context window, aimed squarely at coding, 3D creation, financial research, and long-running agentic work.

Moonshot AI has released Kimi K3.1 Pro, the debut of the K3.1 line and its answer to the question the open-weight community keeps asking: how close can freely downloadable weights get to the best proprietary models? The company’s answer, published on its own blog, is a 3T-class model with a 1-million-token context window, tool calling, and a reasoning-effort control that lets developers trade speed against depth.

Two variants ship together. The first is the flagship; Kimi K3.1 Pro Highspeed delivers two to three times its speed for interactive sessions where latency matters more than raw power. Both are live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API, with the full weights released on Hugging Face by Nov 4.

The split matters more than the naming suggests. A single model tuned for everything tends to disappoint in both directions, so Moonshot tuned one checkpoint for capability and one for response speed, and exposes both through the API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Zero Data Retention is available for API customers, and Kimi Enterprise administrators can switch the model on for their organization; access is off by default at launch.

Third-party tooling picked the model up almost immediately, with the model appearing inside developer environments such as Cursor alongside Moonshot’s own agent products – a sign the weights arrived complete and integration-ready rather than as a teaser.

One detail of the release deserves more attention than it has received: Moonshot describes the model as 3T-class, which in practice means a mixture-of-experts system whose total capacity sits at the trillion-parameter scale while individual requests activate only a fraction of it. That distinction is what allows a free tier and a hosted API to coexist with weights large enough to challenge hosted proprietary services.

Kimi K3.1 Pro is a post-trained version of Kimi K3, and the difference between them is mostly compute applied in the right places. Moonshot says it ran substantially larger post-training, scaling reinforcement learning jointly across hundreds of downstream domains rather than polishing a handful of headline tasks.

That only works if the machinery underneath holds. The company says it optimized every layer of the stack: the RL framework itself, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared platform for checking environment quality. The design goal was blunt – each additional unit of RL compute should buy a stable, predictable gain – and Moonshot claims consistent improvements in every domain it tracks.

Reinforcement learning at this scale fails in boring ways: environments that silently reward the wrong behavior, rollouts queuing behind slower ones, context that fragments as trajectories grow long. A shared environment-quality platform sounds administrative and is actually strategic, because one poisoned environment can flatten returns across hundreds of tasks. Moonshot even notes that cyber capability rose so steadily with post-training scale that it had to intervene – a story covered below.

For readers comparing leaderboards: the Kimi K3.1 Pro numbers in the announcement were produced with reasoning effort set to max, temperature 1.0, and top-p 0.95, so the published scores assume the model is allowed to think as hard as its budget permits.

The context window is the other headline specification. Kimi K3.1 Pro ships with one million tokens, enough for an agent to hold an entire mid-sized repository, its issue history, and a long running log in active view. For coding agents, context of that scale changes failure modes: instead of guessing at interfaces it cannot see, the model can read the definition before changing the call site.

Moonshot positions Kimi K3.1 Pro for end-to-end financial workflows, and the headline claim is a state-of-the-art 64.4 on OfficeQA Pro v2, a benchmark that probes complex fiscal analysis over real financial documents through retrieval, cross-document verification, and multi-step numerical reasoning. On Frontier Finance, measuring support for complete investment workflows from screening to profitability analysis, Moonshot reports a clear improvement over K3.

The internal numbers are larger than the public ones. On Moonshot’s own Financial Modeling test – formula linkage, financial logic, final workbook quality – Kimi K3.1 Pro scores 57.6 against K3’s 50.4. On Internal Financial Research it reaches 70.4 versus 58.8. The most dramatic jump is Internal Business Operation, a long-horizon management simulation, where it achieves about 4.2 times K3’s performance: 37 against 8.9.

Investment research demands factual accuracy, so Moonshot stress-tested hallucination on Kimi Hallu Bench Hard, an internal set built to provoke errors. Claim-level hallucination fell from 23.4% for K3 to 8.3% for Kimi K3.1 Pro, and the share of responses containing at least one factual error dropped to 42.4% of the predecessor’s rate. Moonshot is careful to note these are stress-test rates, not everyday error rates.

On the same bench without tools, the published spread runs from GPT-6 Astra at 4.2 and Kimi K3.1 Pro at 5.4, through Claude Opus 5.5 at 6.8, Claude Fable 5.1 at 11.4, DeepSeek-V4.1-Flash-1017 at 13.1, to GLM-5.3 at 14.8 – putting the new model near the top of the factual-accuracy table among current releases.

Moonshot has form here: when its earlier Kimi model drew scrutiny for hacking capability, the company publicly reduced cyber capability in response – a policy we covered at the time. For Kimi K3.1 Pro it went further and published the arithmetic of that choice.

During development, vulnerability-research performance climbed steadily with post-training scale. Tracing the cause, Moonshot found cyber-related content embedded in the general mid-training corpus and cyber tasks inside the RL environment mix – a natural byproduct of training on broad data. Internal experiments showed that without removing that content at either stage, downstream cyber capability would rise substantially. So for the open-weight release, the company identified and removed cyber-related data from both training stages while keeping general-purpose coding tasks intact.

The effect is measured on xdev-bench-hard, Moonshot’s in-house benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks reproducing recently disclosed bugs in the Linux kernel, Node.js, JavaScriptCore, Exim, and VirGL with mitigations enabled. Each task hands the agent vulnerable source, a short bug hint, and a binary; success means building an exploit, achieving arbitrary code execution in a sandbox, and retrieving a secret string.

On pass@3, the checkpoint trained without any risk-reduction measures solves 90.5%. The final Kimi K3.1 Pro reaches 81.0% – still a large gain over K3’s 33.3%, and behind GPT-5.6 Sol’s third-party-evaluated 100, DeepSeek-V4.1-Flash at 85.7, and ahead of GLM-5.3 at 57.1. Moonshot states plainly that its offensive cyber capability sits below what its own pipeline can produce and behind frontier proprietary models publicly available for at least three months.

The stated principle is a deliberate middle path: never release open weights that extend the absolute frontier of cyber capability, but progressively raise the defensive capability available in open weights, because teams auditing legacy code they maintain benefit when strong tooling runs on infrastructure they control.

The rollout landed the day of announcement across the whole Kimi product family. Consumer users get it on kimi.ai and the iOS, Android, and HarmonyOS apps; desktop users get Kimi Work version 3.1.0 or later for Windows and Apple silicon Macs; developers select kimi-k3.1-pro inside Kimi Code with the /model command; API users pick kimi-k3.1-pro or kimi-k3.1-pro-highspeed on the Kimi API platform, with Zero Data Retention available.

Full model weights followed by Nov 4 on Hugging Face under the moonshotai organization, continuing the company’s practice of publishing the checkpoints that power its consumer products. Kimi Business adds enterprise-grade data privacy and member management with complete separation between personal and organization accounts, and enterprise administrators must enable the new model explicitly.

The distribution is freemium at the consumer tier and metered by tokens through the API, with the Highspeed variant priced as the premium lane for latency-sensitive agents. Because the weights are open, hosting choices multiply quickly: community providers and inference platforms typically list a model within days, which historically compresses effective prices toward self-hosted cost – the same dynamic that made earlier Chinese open-weight releases the default for cost-sensitive workloads.

One caution for evaluators: Moonshot’s blog numbers were run on internal harnesses with max reasoning effort and generous token budgets, and several comparisons borrow scores from other vendors’ blogs. Independent leaderboards will take a week or two to normalize the picture.

Self-hosters should budget accordingly. A 3T-class model will run best on multi-node hardware, so most teams will rent capacity from inference providers rather than buy it, and the economics favor the Highspeed variant for interactive traffic and the flagship for overnight batch jobs. The open weights keep that negotiation honest: any provider drifting on price can be traded for self-hosted capacity.

Placed in context, Kimi K3.1 Pro is the strongest statement yet from the Chinese open-weight camp in a year that has seen the field turn into a genuine arms race. Mistral’s Le Chonk claimed the open-weight crown in late October with a trillion-parameter class release aimed at Europe’s sovereignty agenda, Meta re-entered with an open model pitched directly against Chinese competition, and Moonshot’s own K3 topped open-weight boards over the summer.

What distinguishes this release is strategy as much as score. Moonshot is shipping a two-speed product line, a consumer 3D studio, and finance-grade accuracy claims, then publishing the weights and a documented cyber-reduction methodology in the same breath. That combination forces rivals to answer a harder question than benchmark tables: whether their next open release will be as transparent about what it removed.

Open weights change who captures the margin from a model this strong. When weights are free, the value shifts to distribution, tooling, and trust – the pattern visible across the open-source AI economy, where hosting platforms and tool vendors monetize adoption while model labs compete for the API and enterprise tiers. The dual-track release (consumer apps, API, weights) is a deliberate play across all three layers at once.

For builders, the practical takeaway is simple: a 3T-class open-weight model with a 1M-token window, credible terminal-agent scores, and a fast variant now exists. Teams choosing between proprietary APIs and self-hosted agents in 2027 will price this release into their roadmaps from day one.

The pattern to watch through 2027 is release cadence. Moonshot shipped K3, then K3.1, and labelled the new line explicitly, which reads like a platform strategy: annual major families with interim releases that compound tool compatibility. Rivals that treat each launch as a standalone event will feel pressure to match the pace or cede the developer mindshare that compounds quietly between announcements.