ModelBrief · Kimi K3.1 Pro · Topic Page
Last updated Nov 20, 2026

Kimi K3.1 Pro Launch, in Full: API on Oct 23, Weights on Nov 5

ModelBrief · Published 2026-11-16 · Fact-check 2026-11-16 · Reference: Moonshot tech blog, Hugging Face model card

Kimi K3.1 Pro launched in two acts: the API and products on October 23, the weights on November 5. Twelve days separate the two acts — and until the promise of act one ("full model weights will be released by Nov 4") was kept in act two, "open" was a claim awaiting verification. This article assembles everything confirmed across both acts, and states what the model is: a post-trained product of Kimi K3, not a new base model.

Act one: October 23, API and products go live

On October 23, 2026 (US; October 24 Beijing time), Moonshot AI published the tech blog "Introducing Kimi K3.1 Pro," and the model landed the same day on Kimi.com, the mobile apps, desktop Kimi Work, Kimi Code and the Kimi API, as kimi-k3.1-pro, alongside a Highspeed variant, kimi-k3.1-pro-highspeed. The blog also made a promise: full weights by November 4. Independent scores arrived on day one: Artificial Analysis scored 55 on Index v4.3.2, No. 4 of 215 models; Simon Willison tested it on OpenRouter the same day, calling it the "first time a soon-to-be open weight model has scored this highly"; Hacker News front page, 504 points. Note the timing: the model was already a usable product on October 23 — "released November 5" is true only of the weights, not of the model itself.

Act two: November 5, the weights go public

On November 2, 04:13:51 UTC, the Hugging Face repo moonshotai/Kimi-K3.1-Pro was created; on November 4, 23:07 UTC (November 5, 00:00 Beijing time), @Kimi_Moonshot announced the weights release: 96 safetensor shards, ~1.56TB (1,561,355,732,732 bytes), MXFP4 weights + MXFP8 activations, quantization-aware training, public ungated download; the ModelScope mirror went live the same day (96/96 shards). As of November 16, Hugging Face shows 46,766 downloads and 2,098 likes. The weights also ship in an 8-bit compressed-tensors version; there is no GGUF as of November 4. The 12 days between the acts are the key to this topic's date anchor: the topic is dated to the weights release (November 5), but the API date (October 23) must be stated alongside it.

What the model is: a post-trained product of Kimi K3

K3.1 Pro is not a new base model. It is a post-trained product of Kimi K3 (the 2.8T base released in July 2026): the same 2.8T total / 104B active sparse MoE, 93 layers (69 Kimi Delta Attention + 24 Gated DeepSeek Sparse Attention), substantially larger post-training compute, joint RL across hundreds of domains. The only architectural change is 24 Gated DSA layers replacing K3's full-attention Gated MLA layers — beyond that, the architecture matches K3. Context is 1,048,576 tokens; input is text, image and video (video input per the model card; the blog lists text and image only); output is text only; reasoning effort is low / high / max, default max. This "post-trained product" framing matters: it explains why K3.1 Pro beats K3 across the board yet leads it by only 0.6 points on LMArena (see the benchmarks article) — the post-training gains concentrate on agent and finance tasks, not general chat.

The full channel picture, two weeks in

Within two weeks of launch, K3.1 Pro's channels were all open: Kimi.com and the mobile apps (chat entry point), desktop Kimi Work (3.1.0+), Kimi Code (switch with the /model command), the Kimi API (platform.kimi.ai, OpenAI- and Anthropic-compatible, optional Zero Data Retention, $1 minimum top-up), OpenRouter (moonshotai/kimi-k3.1-pro, listed Oct 23), Amazon Bedrock (Highspeed, model ID moonshotai.kimi-k3.1, announced Nov 6), Hugging Face weights (public download) and the ModelScope mirror. Enterprise access is off by default and must be enabled by an admin. One transition arrangement is worth noting: Kimi K3 (kimi-k3) is served by K3.1 Pro until December 7, 2026 — anyone using K3 before that date is actually running K3.1 Pro.

Moonshot background: valuation and IPO are reported, not confirmed

The developer is Moonshot AI, Beijing. Chinese reporting (aisidaily, Oct 24) puts the company's valuation at about $50B in early October, with a Hong Kong IPO planned early next year — both are marked "reported," not confirmed, and this site does not cite them as fact. What is confirmed: Kimi is one of China's largest open-weights labs, and K3.1 Pro is this site's first Kimi topic, closing a coverage gap.

Day-one reaction: hot in the specialist press, cold in the mainstream

Day-one reaction split two ways: the specialist AI world was hot — AA scored it the same day, Simon Willison tested it the same day, Hacker News front page at 504 points, OrcaRouter published a benchmark-by-benchmark source audit on October 24; the general mainstream was cold — as of November 16, no TechCrunch, The Verge or Reuters piece has surfaced, and in Chinese no dedicated 机器之心 or 新智元 feature, only about eight technical outlets including Phoenix Tech and 掘金. That split is itself the K3.1 Pro profile: evidence depth is sufficient, reach is not.

FAQ

When was Kimi K3.1 Pro released?

Two dates: the API and products launched October 23, 2026 (Oct 24 Beijing time), and the full weights went public November 5, 2026 Beijing time (Nov 4 US). This topic is dated to the weights release, with both dates stated.

Was it an open model before the weights release?

No. Between October 23 and November 4 it was an API-only model, with "coming soon" open weights as a promise in the blog; only on November 5, with the weights public and ungated, did it become a genuinely downloadable open-weights model.

Was Highspeed available at launch?

Yes. Highspeed (kimi-k3.1-pro-highspeed) launched the same day as the standard tier, October 23, and Bedrock serves Highspeed from November 6. But "2–3× faster" is a vendor claim with no published tok/s figure to date, and there is no standalone Highspeed weights repo.

ModelBrief's take: the two-act launch is the easiest thing to get wrong about K3.1 Pro — writing only November 5 makes the model 12 days younger than it is, writing only October 23 hides the weights release, the real open event. The correct form is both dates side by side: October 23 is the product launch, November 5 is the open promise kept. And the model's own positioning needs the same precision: not a new base model, but a post-trained product of K3 — the framing that decides all of its strengths and weak spots.

Reference: Moonshot — kimi.com tech blog, "Introducing Kimi K3.1 Pro" · Hugging Face — moonshotai/Kimi-K3.1-Pro model card · OpenRouter — moonshotai/kimi-k3.1-pro (rewritten and organized by ModelBrief; facts per the official blog and model card)