Moonshot releases Kimi K3.1 Pro: a 3T-class open-weight flagship with 1M-token context

Moonshot AI has launched Kimi K3.1 Pro, the first model in its new K3.1 series and what the company calls its most capable 3T-class open-weight model to date. The model ships with a 1-million-token context window, native image input, and a Highspeed variant that runs at roughly 2-3x the output speed. It is live today on Vercel AI Gateway, and Moonshot has committed to releasing the full model weights by November 4.

For agent builders, this is a rare combination: a frontier-adjacent flagship that is commercially available through a gateway from day one, with open weights on a dated runway rather than a vague "soon."

A post-trained K3, scaled up with RL

Moonshot's official quickstart describes K3.1 Pro as a post-trained version of Kimi K3 with substantially larger post-training compute. To run reinforcement learning jointly across hundreds of downstream domains, the company says it reworked the RL framework, long-context inference, training efficiency, sandbox infrastructure, and rollout scheduling so that each increment of RL compute produces stable gains.

The capability pitch centers on four areas: coding, 3D creation, financial research and modeling, and long-horizon agentic work. In one internal test Moonshot published, the model iteratively corrected 36 errors in shoulder-muscle moment-arm calculations to get an anatomically simulated human walking continuously for 11.2 seconds, up from collapsing within seconds. That anecdote is vendor-reported, not independently reproduced, but it shows the diagnose-and-retry loop K3.1 Pro is trained for.

What Moonshot claims it is good at

Moonshot states that K3.1 Pro achieves state-of-the-art results on OfficeQA Pro v2 and SpreadsheetBench v2, the first testing complex fiscal analysis over real financial documents, the second end-to-end spreadsheet workflows including formula dependencies and error diagnosis. On the company's internal finance benchmarks, it reports a 7.2-point gain over K3 on Financial Modeling, an 11.6-point gain on Financial Research, and about 4.2x K3's performance on a simulated business-operations task.

One caveat worth noting: the official quickstart page currently shows unfinalized placeholder text in several places (benchmark figures appear as "[xx.x%]"), and Vercel's changelog mentions only that the model "makes fewer factual errors than earlier Kimi models on difficult prompts" without numbers. Treat the named benchmark wins as directionally real but not yet citable scores, and check the page again once Moonshot fills in the evaluation tables or ships the weights.

Availability today: one API key on Vercel AI Gateway

Vercel confirms both variants are callable now through AI Gateway using the model names moonshotai/kimi-k3.1-pro and moonshotai/kimi-k3.1-pro-highspeed. Access runs through the AI SDK, the Chat Completions, Responses, and Anthropic Messages APIs, and Moonshot's platform exposes the same model via OpenAI- and Anthropic-compatible endpoints with reasoning effort set by a top-level reasoning_effort field (low, high, max; default max).

Vercel also routes the models to coding agents: install the latest Vercel CLI, run its AI Gateway setup, and select either Kimi model inside Claude Code, Codex, or Cursor. On pricing, Vercel says AI Gateway reflects provider pricing with no markup and charges no platform fee on inference, including bring-your-own-key requests. The changelog adds gateway-side features - Zero Data Retention support, spend tracking, budgets per API key, request traces, and routing rules - and notes Moonshot plans to add more providers once the weights ship.

The open-weight deadline: November 4

Moonshot says full model weights will be released by November 4. That makes K3.1 Pro a lifecycle story in three stages: API availability now (confirmed by two independent organizations), a Highspeed variant for latency-sensitive agents, and open-weight publication later this year. Once weights land, self-hosting, licensing terms, and per-token economics on open infrastructure become the real comparison points against closed flagships.

Until then, the practical route is the commercial API. If your agent workload is long-context coding, 3D or asset-generation scripts, or financial modeling with linked spreadsheets, K3.1 Pro's launch framing targets exactly that lane, and the Highspeed variant at 2-3x output speed addresses the usual complaint about big reasoning models in interactive loops.

Worth bookmarking: the model IDs, the November 4 weight date, and Moonshot's benchmark page before you wire it into production agents. If you are comparing gateways and agent stacks, our guide at /guides/ai-agent-tools covers the current landscape.

Sources and caveats

Primary claim: Moonshot's official K3.1 Pro quickstart (3T-class post-trained open-weight flagship, 1M-token context, Highspeed variant, state-of-the-art on named finance benchmarks, weights by Nov 4). Corroboration: Vercel's official changelog independently confirms the K3-on-K3 lineage, 1M-token context with image input, the November 4 weight date, and live AI Gateway availability with no-markup provider pricing. Both pages are official vendor statements; neither includes independent third-party benchmark verification, and the placeholder figures on Moonshot's page are flagged above.