K3.1 Pro weights are public at the wire: the repo was ready on November 2

The October 23 tech blog wrote the same sentence twice: “Full model weights will be released by Nov 4, 2026.” Over the next two days this site ran public-listing checks on the mornings of November 3 and 4; the newest base repo under moonshotai was still Kimi-K3 (September 2) both times — and “deadline passed, no release” went into the November 4 articles and the catalog. Now the record continues: the first public evidence appeared at 15:59Z on November 4 — 23:59 Beijing time, the final minute of November 4. A full 96-shard mirror (AleaiactaEst1/Kimi-K3.1-Pro, file list verified) was created at that moment; a mirror requires having the weights first, which is indirect evidence of public visibility.

Then comes the mirror cluster: new mirrors at 16:19, 16:25, 16:30, 16:40 and 17:00 (ml-resources, LithosAI, ojuliio, AMAIMedia, AlephFunk expert slices), with unsloth's full mirror at 18:38Z. This site's 3-hour collector first spotted the official repo at rank 5 in HF's open-source-growth trending lane at 19:03Z on November 4 — already 03:03 Beijing on November 5 — observing 797 downloads and 356 likes at the time. The morning re-check on November 5: all 96 safetensors shards present, totalling 1,561,355,966,963 bytes ≈ 1.56TB, single commit stamped 11-02T04:13:51Z. Read in UTC or in Beijing time, “by November 4” was honored at the wire: the first public evidence (the 15:59Z mirror) falls inside November 4 in both zones, while this site's direct sighting came in the early hours of the next day.

Two details make this “at the wire” worth recording. First, the official blog still carried the “by Nov 4” sentence when this site read it on November 5, with no new release announcement — the release itself was quiet, and today's confirmation rests on the repo alone. Second, put the November 4 morning “deadline passed, no release” next to today's record: that sentence was accurate for its moment of checking, but it cannot describe what happened late that same night. A small lesson of promise-tracking: a conclusion at check time is not automatically the ending.

The official repo carries a single commit: “Initial commit”, stamped 2026-11-02T04:13:51Z, with the repo created the same instant. In other words, the 96 shards, README, model card and license were all committed two days before the promise deadline — the content was ready, and the publication was scheduled to the last moment. Why the late night of November 4? The official account doesn't say, and this site won't guess; what can be verified is HF's behavioral boundary: the public listing had no such repo on the morning of Nov 4 (08:35 Beijing), and the first mirror evidence appeared at 23:59. When a private repo goes public, HF keeps the original createdAt, so “created Nov 2, visible late” is compatible with HF's mechanics; but the private period cannot be confirmed from outside, and both ends of the evidence are recorded as they are.

The distinction matters for readers: to judge whether weights are actually usable, the repo's visibility is the only checkable fact; the content stamp only says how ready it was. The two checks on the morning of November 4 were not wrong — at their moments the public listing truly had no weights; what would be wrong is treating a check-time conclusion as the end of the story. The catalog line “the deadline has passed with no release” has been corrected this round.

The README summary: a 2.8T-total / 104B-active MoE (93 layers, 896 experts, 16 selected plus 2 shared per token), 1M-token context; K3's full-attention Gated MLA replaced with KDA + Gated DSA sparse attention; natively multimodal with the MoonViT-V2 vision encoder (401M), taking text and image input. The most informative row is quantization: **MXFP4 weights, MXFP8 activations, quantization-aware training** — weights were laid out at MXFP4 during training, not compressed after the fact. That explains the size: 2.8T parameters at MXFP4 (about 0.5 bytes per parameter) plus scales, 96 shards totalling ≈1.56TB, largest shard 16.99GB (aggregated by this site via the blobs API; the page display agrees at 1.56 TB).

The eval table fixes effort max, temperature 1.0, top-p 0.95, with per-row harness notes (most models on Claude Code, GPT series on Codex, Kimi rows on Kimi Code). A few rows show the shape: Terminal-Bench 4.0 at 65.7 (Opus 5.5 66.4, GPT-6 Astra 58.2 — the open side nearly touching the closed flagship); claim-level hallucination (Hallu Bench Hard w/ tools) at 8.3, the lowest among comparisons except Astra's 5.5, with the blog citing K3's same-caliber 23.4→8.3; BenchCAD w/ tools 92.7 closing on Opus 5.5's 96.2. The weak spots are equally clear: ProgramBench (binarized: a task scores 1 only above 95%) at 31.5 far below Opus 5.5's 65.0; HLE-Diamond 39.4 against Astra's 65.8. All vendor figures, nothing reproduced here.

One more thing ties this release to yesterday's experiment article: the official text states the open-weight build had cyber-related data removed at both training stages — “For the open-weight release of K3.1 Pro, we identified and removed cyber-related data from both training stages.” On xdev-bench-hard (internal vulnerability test, pass@3): 90.5% for the unfiltered checkpoint, 81.0% for the release build, 33.3% for K3 — data removal cuts the same training pipeline's score by about 9.5 points, and the official text adds that K3.1 Pro's cyber capability remains behind frontier proprietary models that have been publicly available for at least three months. Willison's weight-reading experiment yesterday used the API version via OpenRouter; whether the weights build is the same artifact as the API one is not stated by the official account, and this site won't guess — what is certain is that anyone can now download the 1.56TB and check for themselves.

The license body is an MIT grant: use, copy, modify, merge, publish, distribute, sublicense, sell, fine-tuning and derivatives are unrestricted. Two numbered conditions carry substance: §2 defines Model as a Service — inference or fine-tuning access giving a third party meaningful control over inputs, parameters or training data — and requires licensees plus affiliates above $20M aggregate revenue over any consecutive 12 months who run MaaS to sign a separate agreement before commercial use; §3 is rarer: commercial products above 100M MAU or $20M monthly revenue must prominently display “Kimi K3.1-Pro” in the UI. §4 exempts internal use and access via official products or certified inference partners. For comparison: Mistral's open-weight license carries a similar MaaS clause, DeepSeek's family is MIT with use-based restrictions, and Moonshot's previous K3 open release (July 27) used a different license text.

For most readers the practical meaning: organizations and individuals below the revenue line can download the 1.56TB and run it themselves. The official deployment routes are ready: vLLM recipes, SGLang cookbook and TokenSpeed recipes, with the platform API (kimi-k3.1-pro) staying OpenAI/Anthropic-compatible. Back to the site's thread: in yesterday's “weights can be read and written”, Willison's weight-reading test via the OpenRouter API used K3.1 Pro as the example of “API-only” access; today it flips into an open-weights example — 3T-class, MXFP4, natively multimodal weights that anyone can download and verify. Tencent's Hy4 official version, same week, still has its “plan to open-source the weights” pending — K3.1 Pro's wire-timing fulfillment gives that shape of promise a new data point: at-the-wire is not a miss, and a check-time conclusion is not the ending.