Moonshot releases Kimi K3.1 Pro: 1M-token context, tops open-weight Terminal-Bench 4.0
Key highlights
Moonshot has released Kimi K3.1 Pro, the first model in its K3.1 series and its most capable 3T-class open-weight model to date, shipping with a 1-million-token context window. In plain terms, it pairs "read an entire book in one breath and then finish the job coherently" long-context ability with the decision to open the weights, so developers get a full model they can self-host rather than something they can only call on someone else's API.
What happened
K3.1 Pro is a heavily post-trained version of K3 with substantially more post-training compute. To make large-scale joint reinforcement learning pay off stably across hundreds of downstream domains, Moonshot re-optimized the whole stack: the RL framework, long-context inference, training efficiency, sandbox infrastructure, rollout scheduling, and a shared environment-quality-check platform. The payoff is that every increment of RL compute yields steady, predictable returns. K3.1 Pro and a faster K3.1 Pro Highspeed are available today on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with full weights due before November 4.
Technical details
On coding, agentic, and knowledge benchmarks, K3.1 Pro matches or approaches the strongest proprietary models. Terminal-Bench 4.0, which tests realistic resource-constrained terminal tasks, gives K3.1 Pro 65.7 percent, ahead of GPT-6 Astra at 58.2 and just behind Claude Opus 5.5 at 66.4. On Moonshot's own Kimi Code Bench 3.0 long-horizon benchmark it scores 61.2 percent versus K3's 55.95. On long-horizon coding, it once built a human walking-dynamics simulator on the open-source human-atlas anatomical model; across 167 missions and 420 independent reviews it located and fixed 36 errors in shoulder-muscle moment-arm calculations, taking the simulated human from collapsing within seconds to walking continuously with alternating strides for 11.2 seconds.
How it compares
Like DeepSeek, Qwen, and GLM, Kimi follows the open-weight route, but this release's hook is the combination of 3T-class scale, a million-token context, and heavily intensified post-training. The long context gives it a clear edge in long-document analysis, whole-repository code understanding, and very long conversations; opening the weights distinguishes it from subscription-only flagships. For enterprises that want frontier capability without being locked into someone's API and billing, K3.1 Pro is a new heavyweight option.
Industry impact
K3.1 Pro is aimed not just at chat but at coding agents, 3D content creation, and financial research — work that genuinely consumes context and reasoning. On games and front-end work it can already combine engines like Unity, Unreal, and Godot with web search and 3D generation to produce complete interactive, animated experiences end to end. Enterprise access is off by default and administrator-enabled, with zero-data-retention for API customers, signaling that Moonshot is also taking to-B compliance and privacy seriously. Once the weights open in early November, K3.1 Pro is likely to become one of the mainstream base models for Chinese developers building their own strong agents.
A few practical notes for anyone planning to use it. First, the weights are not available yet — they land by November 4, so today the realistic path is the hosted Kimi API or OpenRouter, and self-hosting becomes an option only after the drop. Second, the million-token context is genuinely usable, not a marketing number: you can push an entire codebase or a long research PDF into one prompt, but you will pay for that context in latency and cost, so reserve it for tasks that actually need it. Third, the Highspeed variant trades some peak quality for two to three times the speed, which makes it the better default for interactive product features where users feel every delay. Fourth, enterprise access is off by default and zero-data-retention is offered, which matters if you are handling customer data and cannot have it used for training. Finally, because this is an open-weight model, the community will soon publish quantization, fine-tune, and serving recipes; if you are waiting to self-host, it is worth tracking those to cut your inference bill. Taken together, K3.1 Pro is less a single product than a platform move: Moonshot is trying to make "strong, long-context, open-weight" the default expectation for a Chinese flagship, and the rest of the field will have to respond.
One more thing worth flagging: because K3.1 Pro is an open-weight release, the weeks after November 4 will likely see a burst of community tooling — quantized versions for consumer GPUs, LoRA fine-tunes for specific domains, and ready-made vLLM and SGLang serving configs. Teams that want to self-host should wait for those to mature rather than wrestling with raw weights on day one, since the ecosystem typically compresses both the hardware requirement and the operational complexity within a month. That is exactly the dynamic that makes open-weight releases strategically different from API-only launches: the model's real-world cost and capability keep improving after release, driven by people outside Moonshot.
Closing perspective
For developers weighing whether to adopt K3.1 Pro, the one-million-token context window is the headline feature, but the open-weight commitment is what gives it lasting strategic weight. A long context is only useful if the cost of filling it is acceptable, and an open release invites the community to drive that cost down through better quantization, smarter caching and leaner serving stacks. Watch the first month of community releases rather than the launch-day benchmark numbers: that is when the practical picture — real inference cost, real quality on long documents, and real hardware requirements — finally comes into focus. If those tools land as expected, this release could do for very-long-context open models what earlier open weights did for the mid-size category.