On October 23, 2026, Beijing-based Moonshot AI introduced Kimi K3.1 Pro, the first model in its new K3.1 series and, in the company's words, its most capable "3T-class open-weight model" to date. The launch was reported widely on October 24, and it matters for one clear reason: Moonshot plans to release the model in full — weights included — by November 4, 2026, so developers can run, study and adapt it on their own machines, not only through Moonshot's servers.
That distinction, between using a model and owning a copy of it, is the heart of this story. Closed frontier models are powerful, but they live behind someone else's pricing and rules. An open-weight model shifts some control to the user, and Kimi K3.1 Pro raises the bar for what an open model can do.
What Moonshot actually launched
Kimi K3.1 Pro is a post-trained version of Kimi K3, released in July 2026. Moonshot did not start from scratch: it took the existing K3 base and put it through a much larger round of further training, using, in the company's description, significantly greater computing resources and reinforcement learning across hundreds of tasks. In reinforcement learning, a model is rewarded for outputs that work — code that runs, tasks completed correctly — and steered away from outputs that do not.
The model has about 2.8 trillion total parameters, with 104 billion active for any given token, the small chunk of text the model processes at a time. It also offers a context window of up to one million tokens, so it can take in a very large amount of text, code or documents in a single request. A faster Highspeed version, said to run two to three times faster at twice the API price, launched alongside it. Both are already available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. The downloadable weights, however, are not out yet — that release is promised by November 4.
Why "open-weight" matters more than "open-source"
People often use "open-source" and "open-weight" as if they mean the same thing. They do not. Weights are the learned numbers inside the model. Publishing them lets a skilled team download the model, run it on its own servers, fine-tune it on private data and inspect its behaviour without sending every request to an outside company. That can help with privacy, cost control and independence: if a provider changes its prices or policies, a team with the weights still has a working model.
Weights alone, though, are not the full recipe. Training data, training code and detailed methods are usually not included, so "open-weight" does not mean anyone can exactly reproduce the model from scratch. Treat Kimi K3.1 Pro as a powerful shared tool, not a fully transparent public project. The real test comes after November 4: does the release arrive on time, under what licence, and can independent teams reproduce Moonshot's results?
What the benchmarks say — and what they do not
On Terminal-Bench 4.0, a test of real terminal-based coding work, Moonshot reports 65.7 percent for Kimi K3.1 Pro, against 66.4 percent for Claude Opus 5.5 and 58.2 percent for GPT-6 Astra in the same comparison. On BenchCAD, a test of writing code that describes 3D geometry, it scored 92.7 percent, within four percentage points of its leading rivals. Moonshot also says it tops its own OfficeQA Pro v2 office-work test, and that on an internal hallucination stress test, the claim-level error rate fell from 23.4 percent for Kimi K3 to 8.3 percent for K3.1 Pro.
Independent evidence points the same way, with a note of caution. Artificial Analysis, an independent benchmarking service, gives Kimi K3.1 Pro an Intelligence Index score of 55, fourth overall and behind only closed models, and 11 points above Kimi K3. That is a meaningful external check, but it is still one index, measured on hosted access before the weights are public. Benchmarks are a signal, not a verdict: they measure fixed tests under fixed conditions, and they do not guarantee equal strength on your code or documents.
That caution applies to every model. When we covered OpenAI's GPT-6 intelligent interface, the story was also about how a model behaves in real use, not a single score. The honest summary here is the same: headline numbers are company-reported, one independent index supports the general picture, and the full independent test is only possible once the weights are out.
Coding, 3D creation and office work
Moonshot pitches Kimi K3.1 Pro most clearly at builders. It says the model covers the full range of coding work, from quick fixes in real codebases to full-project builds, and can deliver complete financial workflows end to end. In its own demonstrations it generated animated simulations of human motion and built multiple 3D video games — a hard test, because a game contains so much code that a single bug can break it.
The 3D angle is worth pausing on. A 3D model can be drawn by hand, or described by code that defines its shape in numbers. Kimi K3.1 Pro takes the second path, turning a description into geometry through code. If that works reliably outside demonstrations, game developers, product designers and educators could prototype faster, starting from a prompt-generated model that a human then refines.
The office-work claim matters for a different audience. Many businesses need a model that can read a long report, pull out the right figures, draft a summary and keep a chain of work consistent across many steps. That is agent work: not one answer, but a sequence of actions with tools. The practical question for early users is simple: can it finish a multi-step job without drifting, inventing figures or needing constant rescue?
Price, access and the Highspeed trade-off
For API developers, Moonshot has cut prices against Kimi K3. Input costs 2 dollars per million tokens and output costs 10 dollars per million tokens, with cached input — repeated instructions the model can reuse — at 0.20 dollars per million tokens, about a third cheaper than K3 according to independent coverage. Highspeed costs twice as much, at 4 dollars per million input tokens and 20 dollars per million output tokens, for two to three times the speed.
For a business running thousands of agent tasks a day, the standard-versus-Highspeed choice becomes real money. Use the standard model where accuracy matters more than speed, and reserve Highspeed for live jobs such as interactive coding help, where a user is waiting.
Access also has tiers in the consumer apps: the model requires a Plus membership, while the full one-million-token context requires Pro. Business users should note that administrators must enable the model, and access is off by default at launch. Moonshot also says Zero Data Retention is available for API customers, meaning prompts are not kept for training — worth confirming in writing before sending sensitive data through any API.
The safety trade-off Moonshot made on purpose
The most interesting part of this launch is not a capability but a deliberate removal. Moonshot says it stripped cyberattack-related training data and tasks from the model's development, while keeping general coding work. The reason is linked to the open-weight release: once weights are public, anyone can use the model, and Moonshot does not want the release to advance offensive hacking.
On Moonshot's own internal vulnerability test, the reported score fell from 90.5 percent to 81.0 percent after this removal — a real cost, presented openly as the price of a safer release. But this is a company-run test, not an independent audit, and researchers will want to test the released weights themselves.
This is where the wider platform race shows. When we covered Amazon's Bee Frontier and Eclipse AI systems, the theme was AI moving from answering questions to acting on a user's behalf. Kimi K3.1 Pro points the same way, but from the open side: an agent-capable model a developer can host, adapt and build into their own product.
What this means for developers, businesses and everyday users
For developers, the immediate value is a cheaper route to a near-frontier coding and agent model through an API today, with the option — from November 4, if the release lands — to self-host. Self-hosting is not free or simple: a model this size needs serious hardware. But for teams with privacy requirements, steady high-volume use, or a need to fine-tune on their own data, having the choice matters.
For other businesses, do not switch tools on a benchmark table. Run a small, real trial: give Kimi K3.1 Pro a job you already do, such as cleaning up a report or drafting code for a known project, and compare its output with your current tool, checking facts and code by running them. Watch for the usual failure modes of every large model: confident errors, invented citations and drift over long tasks.
What to watch before November 4
Three checks will tell us whether Kimi K3.1 Pro lives up to its launch. First, November 4 itself: do the full weights arrive on time, with a clear licence that permits the uses developers expect? Second, independent reproduction: once researchers can run the model themselves, do the Terminal-Bench, BenchCAD and office-work results hold up? Third, real-world agent tests: can it complete long, multi-tool jobs reliably, and does the cyber-capability reduction survive fine-tuning attempts in published research?
Moonshot has also not specified the open-weight licence in launch coverage; licence terms decide whether a business can use the model commercially.
The bottom line
Kimi K3.1 Pro is the most significant open-weight launch of October 2026 so far. It takes a model class that was, until recently, clearly behind the closed frontier on coding and agent work, and puts it within touching distance on the tests Moonshot highlighted: 65.7 percent against 66.4 percent on Terminal-Bench, a fourth-place independent index score, and a price cut of about a third. The gap is now small enough that the open-versus-closed choice is becoming a practical business decision about control, privacy, cost and speed, not just which model is smartest.
Try the hosted model now if you build or buy AI tools, judge it on your own work rather than its launch slides, and watch November 4 closely. If the weights arrive as promised and independent tests confirm the headline results, Kimi K3.1 Pro will have made near-frontier AI something a much wider group of developers can actually own, not just rent.