"Kimi K3.1 Pro" is Moonshot AI's newest model word, and this page is built from first-hand captures of Moonshot's own pages: the official English tech blog ("Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence"), the moonshot.cn research feed, and the Kimi API platform, all captured November 1, 2026 observed. Moonshot's framing: K3.1 Pro is "the first model in our new K3.1 series and our most capable 3T-class open-weight model to date, with a 1-million-token context window," delivering "substantial improvements across coding, 3D creation, financial research, and other agentic capabilities" vendor claim. There is also a speed sibling — Kimi K3.1 Pro Highspeed — and a weights promise with a date on it: full model weights released by November 4 vendor claim. Model card, API access, and the vendor-reported numbers, all labeled below. Labels: vendor claim / observed / analysis.
| Field | What the official pages say |
|---|---|
| Name | Kimi K3.1 Pro observed |
| Maker | Moonshot AI observed |
| Announced | Late October 2026 — research feed entry dated October 24, 2026 observed |
| Series | First model in the K3.1 series; post-trained version of Kimi K3 with substantially larger post-training compute vendor claim |
| Scale | 3T-class (trillion-scale) parameters vendor claim |
| Context window | 1 million tokens vendor claim |
| Stated strengths | Coding, 3D creation, financial research, agentic capabilities vendor claim |
| Speed variant | Kimi K3.1 Pro Highspeed — stated 2-3× the speed of K3.1 Pro vendor claim |
| Availability | Kimi.com, Kimi Work, Kimi Code, and the Kimi API (per the blog, available at announcement) vendor claim |
| Open weights | Full model weights to be released by November 4 (vendor's stated plan) vendor claim |
| API pricing | Not in the pages we captured — details await official documentation analysis |
Moonshot's blog publishes a benchmark slate. These are Moonshot-reported figures vendor claim — worth tracking, not worth citing as settled fact. Note that the comparisons are to other vendors' models as measured by Moonshot analysis:
| Benchmark (per Moonshot) | Kimi K3.1 Pro | Vendor-stated comparisons |
|---|---|---|
| Terminal-Bench 4.0 (coding agents on terminal tasks) | 65.7% vendor claim | Ahead of GPT-6 Astra (58.2%), "narrowly behind" Claude Opus 5.5 (66.4%) — all as measured and reported by Moonshot vendor claim |
| Kimi Code Bench 3.0 Long-horizon (Moonshot's internal multi-step coding benchmark) | 61.2% vendor claim | Above Kimi K3's 55.95% vendor claim |
The Terminal-Bench claim is the one that powers the "Kimi vs everyone" discourse this week — including our Kimi K3.1 Pro vs Claude Opus 5.5 page, where we put it in context. One number on one vendor-run benchmark is a data point with an owner, not a verdict analysis.
If you searched for the Kimi K3.1 Pro API, here is what Moonshot's own platform pages state observed: the API platform banner announces "Kimi K3.1 Pro has launched!", and the quickstart describes calling K3.1 Pro through the Chat Completions API, Responses API, or Messages API — create an API key, choose the model, configure base_url, and send requests over plain HTTP or through OpenAI- or Anthropic-compatible SDKs vendor claim. The platform also advertises adjacent capabilities (vision input, context caching, tool calling) for its model lineup generally — per-model support details are a documentation question, not an announcement question analysis.
Most frontier-launch weeks end with an API you cannot audit and weights you cannot download analysis. Moonshot's blog instead attaches a calendar to openness: full weights by November 4 vendor claim. If it lands on schedule, K3.1 Pro becomes one of the largest open-weight models available for self-hosting — which is why the open-weights community is watching this launch more closely than the benchmark table analysis. If it slips, the "open-weight" framing at announcement time will have been a launch promise, not a shipping fact; both outcomes are informative, and only one of them is currently verifiable analysis.
Moonshot's pages we captured do not state API pricing analysis. "Free" claims circulating in launch-week posts are not first-hand sources. The model itself is open-weight (weights promised by November 4), but self-hosting a 3T-class model has real infrastructure costs — free weights are not free compute analysis.
Moonshot reports K3.1 Pro scoring ahead of GPT-6 Astra on Terminal-Bench 4.0 (65.7% vs 58.2%) vendor claim. That is one benchmark, run by the vendor claiming the win, on Astra — not its successor generation analysis. No independent head-to-head exists yet.
1 million tokens, per Moonshot's blog and its homepage positioning vendor claim. Usable-context behavior at that scale (degradation curves, caching) is a testing question for your own workload analysis.
The plan is yes — Moonshot states full model weights will be released by November 4 vendor claim. "Locally" is doing heavy lifting for a 3T-class model: check the released license terms and hardware requirements when weights actually ship analysis.
A companion variant Moonshot says delivers 2-3× the speed of K3.1 Pro, available on the same surfaces (Kimi.com, Kimi Work, Kimi Code, Kimi API) vendor claim. Speed/quality trade-off details are not quantified in what we captured analysis.
This page is descriptive and independent, built from direct captures of Moonshot AI's official tech blog, research feed and API platform, last reviewed November 1, 2026. Not affiliated with Moonshot AI. "Kimi" is a trademark of Moonshot AI; benchmark figures above are vendor-reported and labeled as such.