Kimi K3.1: Four Signals Point at a Launch Next Week, and the One Commit Nobody Can Find

The most load-bearing claim about Kimi K3.1 this week is a code commit that, as far as anyone outside China's tech press can tell, does not exist. On 19 October 2026 several Chinese outlets reported that SGLang — the open-source inference stack — had taken a change titled "Kimi-K3.1 support for the K3 chat encoder," adding a fourth reasoning tier called medium and 107 test cases, and tagging a Hugging Face repository named moonshotai/Kimi-K3.1 at a specific revision. None of those reports linked the commit. I went looking for it in the public history of the SGLang repository for 17–20 October and could not find it, and no pull request with that title is open or closed there today. That does not make the story false. It makes it unverified, which is the correct description of almost everything about the next K​imi model, including the practitioner post that got this article queued: a developer praising Kimi K3 for building an animated plant-care SaaS site, and adding, in passing, that "a new model is cooking, we can see it next week." The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

So this is a status page for a model that has not shipped. What is established is the baseline: Kimi K3, the 2.8-trillion-parameter open-weight flagship Moonshot AI announced on 16 July 2026 and published the full weights for on 27 July, with a 1M-token context window and three reasoning levels. Everything above that baseline — the name, the fourth reasoning tier, the release week, the price, the licence — is signal, and the signals are of wildly different quality. Sorting them is the whole value of a piece written this early.

The four signals, graded

Each of these is a separate trace, and they are not equally strong. Ranking them is more useful than listing them.

• Strongest — the front-end identifier, 17–18 October. Developers inspecting Moonshot's own website files found the string k3d1-agent, alongside markers reading make3d, BOT, AUTOMATION_K3 and OK_COMPUTER_SWARM, and a new bot module whose purpose nobody outside the company knows. This is first-party code on a first-party domain. It is not a model, but it is the vendor's own naming machinery, and it names something that is not K3.

• Strong-ish — the API registry sighting, 28 September. The identifier kimi-k3-1 appeared in Moonshot's backend model registry, reportedly responding to interface probes, and the platform briefly advertised a K3.1 / "K311111" model with a 1M-token context window. Three weeks later there is still no model card, no price and no public identifier you can call through the standard API list.

• Weaker than it reads — the reasoning-tier commit, 19 October. Detailed in its retelling, unlocatable in the source. Treated on its own below.

• Weakest, and the one that queued this piece — a practitioner's aside, 20 October. A developer who likes K3 says a new model is "cooking" and visible "next week." No name, no spec, no screenshot, no claim to inside knowledge. It is enthusiasm with a date attached. That it arrived in the same week as the other three is what makes it worth a line, not the post itself.

None of the four is a release. Moonshot's blog still ends at Kimi K3 and PerceptionBench on 16 July. The platform documentation still lists kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed and kimi-k2.6. The moonshotai organisation on Hugging Face has nothing newer than Kimi-K3, created 13 June 2026 and last modified 2 September. And the vendor's own rate card has four rows, none of them a successor.

What the reported commit actually claims, and why the name would matter

The retellings agree with each other closely, which is worth something even when the primary source is missing. According to them the change adds a chat-encoding adapter for K3.1 to SGLang's experimental router, ships 107 tests covering chat formatting, tool calls and image input, and notes that K3.1's tokenizer files are byte-identical to K3's, reusing the same 16 special tokens. The one functional difference named is the reasoning ladder: K3 offers low, high and max, and K3.1 reportedly inserts medium between low and high, with routing still defaulting to max.

That detail has the ring of truth because it matches what is already public. K3 does not accept temperature, top-p or seed; reasoning depth is set through a reasoning_effort parameter with exactly those three values, according to the platform documentation and our own model page. A successor adding a fourth stop on that ladder is the least dramatic possible change and the easiest to describe accurately from a diff. It is also the kind of change an inference stack would need to encode before a launch rather than after one, because the router has to know how to render a request the model will accept.

The stronger claim in the retelling is the name. A commit that references moonshotai/Kimi-K3.1 at a specific revision implies a repository exists in the vendor's organisation, even if it is private or not yet indexed. That is a different class of artefact from an arena codename or a leaked screenshot: repository identifiers are what weights actually get published under. But it is also the claim I could not check. The Hugging Face API returns 401 for moonshotai/Kimi-K3.1 — as it does for any name that does not resolve to a public repository, including deliberately fake ones, so that test tells us nothing either way.

What can be checked is the substrate. SGLang really has been building out K3 routing: a pull request rendering Kimi-K3 through its Rust processor merged around 13 October, with commits through that week closing render gaps, parsing output and lowering K3 chats onto the engine's generate path. A follow-up that extends that work to a K3.1 encoder is exactly the next commit you would expect. A story that fits the engineering pattern is more plausible than one that does not — and still not evidence that this particular commit exists.

K3 versus the model being described

The baseline matters more than usual here, because if K3.1 is a point release on K3 the way the tokenizer claim suggests, then most of what you would want to know is already answerable from K3.

• Release — K3 announced 16 July 2026, full weights on Hugging Face 27 July 2026; K3.1 has no date, only a reported week.

• Size — K3 is 2.8 trillion total parameters with roughly 104 billion active per token, per Artificial Analysis; K3.1's size is unknown, and any figure attached to it is inference from K3.

• Context — K3 is 1,048,576 tokens; the September registry sighting and the platform teaser both pointed at the same 1M window for K3.1, which would make context parity rather than an upgrade.

• Reasoning — K3 exposes low, high and max; K3.1 is reported to add medium, making four tiers.

• Modality — K3 takes text and image in, text out, natively; nothing reported suggests K3.1 changes that.

• Price — K3 lists at $3.00 per million input and $15.00 per million output tokens, with a 90% cache discount; K3.1 has no price anywhere, and a point release usually inherits rather than cuts.

• Licence and weights — K3 ships under the Kimi K3 License, a custom "other" licence rather than a standard open-source one; whether a successor opens at all is the open question, and it now has a policy dimension.

The honest summary of that comparison is that nobody has published a single K3.1 specification. The effort-level addition is the only functional difference in circulation, and it comes from a commit that cannot be located. Anyone quoting a parameter count, a benchmark or a launch date for K3.1 is extrapolating.

The policy change that decides whether you can download it

This is the part of the story that is not a rumour, and it is the reason a leak piece about a Chinese open-weight flagship is worth reading this particular week. On 16 October 2026 China's Ministry of Commerce and Ministry of Science and Technology published Announcement No. 48 of 2026, a draft revision of the technology export catalogue that, for the first time, reaches AI model weights and source code. The coverage threshold reported across the trade press is two trillion parameters or more, or a place in the global top ten on performance and safety benchmarks. Forty technologies are added in total, 23 of them AI-related, with 14 slated for prohibition and 26 for restriction under licence. Public comment runs to 15 November 2026.

Two carve-outs do most of the work for anyone reading this from outside China. Technologies already in the public domain through open-source release or academic publication are excluded, and API and SaaS access to foreign users is not restricted. One outlet additionally reported that an interim registration measure was operative from 16 October while the catalogue itself remains in consultation; other coverage describes the whole instrument as a draft with comments open, and the two readings have not been reconciled in public. Treat the draft status as established and the interim measure as reported.

Why this lands on K​imi specifically: K3 is over the two-trillion threshold by a wide margin, and it is one of the models named in coverage of the draft as sitting above it. K3's weights are already out, so the public-domain carve-out is the natural reading for K3 itself — that is an inference from the text as reported, not a ruling. The live question is the successor. If K3.1 launches after the catalogue is finalised, the difference between shipping weights on Hugging Face and shipping only an API is no longer a product decision made in a conference room; it is a licensing question with a ministry on one side. A week-of-launch tease and a draft export catalogue landing six days apart is the kind of coincidence that changes what a reader can plan for, and it is the one thing in this story that is both new and sourced.

Practically, that argues for a specific posture: if you are waiting on K3.1 because you want to self-host it, do not build a plan that assumes it. If you are waiting because you want to call it, the API carve-out in the draft is the part that makes that likely regardless of how the weights question resolves.

What you can call today

Kimi K3 is on our catalogue as kimi/kimi-k3: 1,048,576-token context, text and image input, $3.00 per million input and $15.00 per million output tokens at the provider's own list price with 0% markup over it. The coding-tuned member of the line, kimi/kimi-k2.7-code, is $0.95 and $4.00 on a 262K context, and the older kimi/kimi-k2.6 and kimi/kimi-k2.5 sit behind it. All four run through one key against 200+ models.

That matters for a leak piece specifically. The mistake people make with an unconfirmed successor is either to stall a project waiting for it or to bet a production path on it the day it appears. Routing solves the second: put K3 behind the same interface a successor would use, and when an unfamiliar model turns up you can send a slice of traffic at it and keep the fallback chain pointed at something you have already measured. Automatic failover is the difference between trying an unproven model and depending on one. Because we pass provider list prices through without markup, a launch-day price on the vendor's rate card is the price on ours the same day — no repricing cycle to wait out.

How to read the next two weeks

Watch for artefacts in ascending order of how expensive they are for Moonshot to produce, because each one is harder to fake and each one settles more.

• A model identifier in the vendor's public API list. That single string answers whether K3.1 exists as a product, and it usually appears before any announcement.

• A rate-card row. Pricing next to the name means the model is being served and billed, whatever the blog says.

• A repository under moonshotai on Hugging Face. This is the one that settles the numbers and the licence, and for this model the licence field is the first thing to read rather than the last.

• An entry on a neutral leaderboard. Artificial Analysis carries Kimi K3 at 44 on its Intelligence Index, ranked fifth of 121 models, and has no K3.1 page. When that page appears, it is the first number in the story produced by somebody who is not the vendor.

Falsification is cheap to state and worth taking seriously. The name may be wrong: the front-end string was k3d1-agent, the registry string was kimi-k3-1, and the press has settled on K3.1 by convention rather than by announcement. The shape may be wrong: a successor could arrive as K3.2, as a K4 at a different size entirely, or as a post-training refresh under the K3 name with no new repository at all. Or the week may simply pass, as the August window did and as the mid-October window a Nomura-sourced report pointed at on 11 October now largely has.

What to do with all of this is short. If you need a 1M-context open-weight model today, Kimi K3 is callable and measured, with three reasoning levels that cover most of what a fourth would. If you need to know whether the successor is real, do not read the commentary — check the model list, then the rate card, then the repository. And if the thing you actually care about is downloading the weights rather than calling the model, the document to read this week is not a leak at all. It is a draft export catalogue with a comment deadline of 15 November, and it will have more to say about what ships open than any commit anyone claims to have seen.