Kimi K3.1: Subscriptions Are Back, and a Successor Is Reportedly in Post-Training

A subscription queue reopening is a strange thing to get excited about. But the Kimi post that went up on September 19, 2026 is doing two jobs at once, and the second one is the interesting one: it says Moonshot AI has already started a new post-training run for the next model in the Kimi K3 line — a run the poster describes as more focused on computer use, with work also going into the multimodal experience — and that the next model, which the post calls the next Kimi K3.1, will be here before the end of October.

None of that second half is confirmed. There is no Kimi K3.1 model card, no weights, no API id, no price and no date. The only model in this family you can call today is Kimi K3, the 2.8-trillion-parameter open-weight system Moonshot shipped on July 16, 2026. So treat this as what it is: an early signal from a leaker with a real track record, arriving in the same week as one genuinely checkable event.

That checkable event is the reason this is worth a post rather than a shrug. Kimi's consumer subscriptions had been closed since July 20, 2026, when Moonshot paused new sign-ups because K3 demand had run past what its clusters could serve. Two months later they are open again, with a rebuilt tier structure. If you were waiting to buy in, that part is real and it happened within the last few days.

The short version. Subscriptions: back, new tiers, Code now starts one level higher than before. Next model: being post-trained right now, aimed at computer use and multimodality, claimed for before the end of October — unverified, from one account, and not something to plan around. Kimi K3: available now, and still the only thing in this story you can actually run.

What the post says, and what it does not

The signal is a single X post from @chetaslua dated September 19, 2026. It carries four claims, and they are not equally strong:

• The subscription is back. Verifiable, and verified — see the next section.

• A new post-training run has started. Not verifiable by anyone outside Moonshot. Post-training is the stage after pre-training where a model is tuned into a usable product; "a new run started" tells you roughly where a successor sits in the pipeline, and nothing about when it lands.

• It is more focused on computer use, with improvements to the multimodal experience. This is a direction, not a spec. It is also the least surprising claim in the post, for reasons worth spelling out below.

• The next model arrives before the end of October. A date. Dates from leaks are the single most testable thing anyone publishes, and this account's previous Kimi date was wrong — more on that in the timeline.

What is missing is what you would actually want: parameter count, context length, price, licence, benchmark scores, or any indication of whether this is a Kimi K3.1-style point release or something larger. The post does not say, and nothing published since has filled the gap. There is also a naming problem. Leaks have floated "K3.1" since July, prediction markets and gray-test watchers have pointed at a "next K-series model," and reports have variously gestured at K3.1, K3.2 and K4. Those are three different products. This post does not resolve which one it is describing.

The part you can check: subscriptions reopened

Moonshot paused new consumer subscriptions on July 20, 2026, four days after K3 launched, saying demand had approached the limits of its existing clusters and that available compute would be redirected to users who had already paid. New slots were to reopen in batches as capacity was added. That pause lasted about two months.

Subscriptions reopened on September 18, 2026, with the tiers renamed and repriced to the same annual totals as the old ones:

• Go — 468 yuan/year, about 39 yuan a month

• Plus — 948 yuan/year, about 79 yuan a month

• Pro — 1,908 yuan/year, about 159 yuan a month

• Max — 6,708 yuan/year, about 559 yuan a month

Two changes matter more than the prices, which did not move. First, the plan to sell Kimi Code separately was largely withdrawn under user pressure — Plus, Pro and Max all still include it. Second, the cheapest tier no longer does, so the entry price for Code effectively rose from 468 to 948 yuan a year. Community threads also report that weekly limits were removed while quotas were reduced, and that the 1M-token long-context benefit moved up from the old mid-tier plan to Max. Those last two are user commentary rather than an official breakdown, and they are the kind of detail that decides whether a plan is worth buying — check them against your own usage before committing a year.

If you are outside China the picture is different in an important way: the tiers above are the domestic consumer plans, and what you can actually buy, and at what price, depends on region. For API access specifically, none of this gates you — the Kimi API is a separate product from the consumer subscription and was never part of the pause.

Why "computer use and multimodal" is the least surprising claim in the post

Moonshot has been building toward exactly this for a year, which is why the direction is credible even though the timeline is not.

Kimi Work, Moonshot's desktop agent for macOS and Windows, shipped with a Computer Use capability: it reads the screen and handles clicking, scrolling and typing, and it runs in the background rather than taking over your mouse. Underneath it, a browser bridge drives a real logged-in Chrome or Edge session over the DevTools protocol for navigation, clicking and form filling. Around that, the company has built Agent Swarm for up to 300 parallel sub-agents, a goal mode that keeps iterating on an outcome for hours, and a scheduler for recurring jobs. That is a product surface that needs a model good at seeing a screen, deciding an action and checking the result.

K3 itself was already pointed that way. It supports native image input, and its launch material leaned on long-horizon agentic work — reading environment state, calling tools, inspecting results, correcting course — which is computer use in everything but name. Moonshot's own technical reporting describes training tasks that can run to thousands of tool calls while retaining file, application and virtual-machine state. A successor whose post-training run is explicitly weighted toward computer use is the obvious next step, not a pivot.

Multimodality is the same story. The K2.5 generation introduced a "zero-vision" fine-tuning stage — text-only supervision used to keep visual reasoning and tool-calling intact — followed by multimodal reinforcement learning. K3 kept a MoonViT-family vision encoder and native image understanding. "Improving the multimodal experience" is what continuous work on that stack looks like from the outside.

None of this makes the claim true. It makes it plausible, which is a different and much weaker thing. Plausible directions are what leaks are best at; they are also what leaks are worst at dating.

Four dates, three of them wrong

The most useful way to read this post is as the fifth entry in a sequence, because the sequence has a track record:

• July 16, 2026 — Kimi K3 ships. The only confirmed item on this list. 2.8 trillion parameters, Mixture-of-Experts, 1M-token context, native image input, open weights.

• July 26, 2026 — the first K3.1 leak. A single X post listing rumored improvements: faster inference, better token efficiency on long reasoning, steadier coding and agent workflows, fewer wasted reasoning steps, still open-weight. No numbers, no date, and it closed by asking readers whether they thought it would win — the shape of speculation, not a briefing.

• August 18, 2026 — the first date. A follow-up from the same leaker put the next Kimi model against the top proprietary systems and said it was expected within August. August ended with no release. This is the one element of the whole story that has ever been testable, and it failed.

• September 18, 2026 — the teaser. Moonshot's own account published a long digit string and deleted it; the community read it as the decimal expansion of π with the leading "3.1" removed. Vendor-side, second-hand, and unverifiable, because the post is gone.

• September 19, 2026 — this post. Subscriptions confirmed, a post-training run claimed, a new deadline: before the end of October.

There is a pattern here that is worth naming. The direction of these leaks has been consistent for two months — efficiency, reliability, agents, computer use, multimodal — and every date attached to them has slipped. A leak can be right about what is coming and wrong about when, and this one has been exactly that so far. The September 18 teaser complicates the picture in the other direction: a vendor teasing its own successor is a stronger signal than a leaker's bullet list, because a company that teases something has it in some form. But the teaser carried no date either, and the post that follows it here is the leaker's, not Moonshot's.

What is actually confirmed, and what to run today

Kimi K3 is the real model in this story. Released July 16, 2026, with weights and a technical report following on July 27, it is a 2.8-trillion-parameter sparse Mixture-of-Experts system with 896 experts and roughly 104 billion active parameters per token, built on Kimi Delta Attention and Attention Residuals, with a 1M-token context window and native image input. It took the top spot on the independent Frontend Code Arena at launch at a reported ~1,679 Elo, ahead of Claude Fable 5 and GPT-5.6 Sol, and posted a reported 93.5% on GPQA Diamond. Label those as launch-time figures rather than settled ones — the boards and the models around K3 have moved since, and the scores come from Moonshot's own release material.

What has not moved is that K3 is genuinely strong at the thing the leak says the successor is being trained for. Long-horizon agent work with vision in the loop — the model looking at what it produced and correcting — is where it was built to compete, and it is the reason the "computer use" claim in the September 19 post reads as continuity rather than news.

Kimi K3 is available through OrcaRouter as kimi/kimi-k3, alongside Kimi K2.5, Kimi K2.6 and Kimi K2.7 Code, on one OpenAI-compatible endpoint. That matters more than usual when a leak is in the air. OrcaRouter passes provider list price through at 0% markup, so if Moonshot cuts K3's price — which is the realistic shape of a "successor ships" event for anyone already building on it — the new price is live on our side the same day, with no contract to renegotiate. And when a model you cannot yet verify does arrive, failover is how you try it without betting a production path on it: route a slice of real traffic, keep the proven model as the fallback, and let the router move requests back if the new one misbehaves.

How this piece will look if it is wrong

The honest test of a leak write-up is whether it named the things that would falsify it. So, concretely, here is what would settle this — and what would kill it.

• A Hugging Face repository or a model card from Moonshot is the only hard confirmation. Nothing else in this pipeline is checkable by a reader.

• A new model id in the Kimi API's own model list would confirm a release even before weights land.

• An anonymous entry in a blind arena is a weak instrument and should be treated as one. Earlier this year much of the community was convinced an anonymous Code Arena codename was the next Kimi model under gray test, on the strength of Moonshot's naming pattern for prior releases; a tokenizer fingerprint later pointed the same entry at a different model family entirely. An arena label is a guess about identity, not evidence of one.

• Silence through October would make this the second consecutive missed date from the same account, and would move the reasonable expectation for a successor into next year — which is not the same as the model not existing.

If you are deciding something this week, decide it on the subscription news, which is real, and on Kimi K3, which is shipping. The successor is a direction with a deadline attached to it by someone who does not work at Moonshot, and the correct amount of planning to do around it is none.

Questions people are actually asking

Is Kimi K3.1 released?

No. As of September 20, 2026 there is no model card, no weights, no API id and no price for anything called Kimi K3.1. The only shipped model in the line is Kimi K3, from July 16, 2026.

Should I wait for it before starting a build?

No — and the subscription news is the reason. If you need K3-class capability now, it is available today and the price has not gone up. If you want to be first on a successor, the low-cost way to prepare is to build against an endpoint that can route to a new model id the day it appears, rather than rewriting an integration around a model that has no published API surface yet.

Does the reopened subscription include API access?

No. The Go, Plus, Pro and Max tiers are consumer plans for Kimi's web, app and agent products, and for Kimi Code on the top three tiers. API access is billed separately, per token, and was never part of the July pause.

What does "more focused on computer use" mean in practice?

It means post-training weighted toward screen-reading and GUI action — the model looking at an interface, deciding what to click or type, and checking whether the result was right. Moonshot already ships that capability in Kimi Work and already trains K3 on long tool-call runs with retained application and virtual-machine state, so a successor tuned further in that direction is the expected next step rather than a new product line.

The next real datapoint is not another post. It is a repository, a model card, or a model id. Until one of those appears, everything past the subscription line in that September 19 post is a claim with a date attached to it — and the date is the part this account has already gotten wrong once.