Kimi Explained: Moonshot AI’s Long-Context Chatbot and the K3 Models
Kimi is not one model. It is Moonshot AI’s whole AI family — a chat assistant, a phone and desktop app, and a line of large language models that, as of 18 November 2026, has a brand-new flagship: Kimi K3.1 Pro, launched on 23 October 2026, with its open weights published on 4 November 2026. The model most explainers still describe as the current Kimi — the K2 series — is being retired: closed to new API users on 31 October 2026, with a full retirement planned by the end of December 2026. That split — a new flagship on top, a retiring line underneath — is the single most important thing to understand about Kimi in November 2026, and it is the spine of this guide. Checked 18 November 2026 against Moonshot’s own K3.1 Pro tech blog, which the researcher read in full, and the Kimi API pricing pages as read by multiple consistent secondary sources.
There is a second confusion to clear up before anything else. “Kimi” is three things at once: the name of the assistant you chat with at kimi.com, the name of the model family behind it, and the name of the individual models inside the family — Kimi K2, Kimi K3, Kimi K3.1 Pro. On top of that, the version numbers do not mean what version numbers usually mean. K2 is not one model but a whole series — K2, K2 Thinking, K2.5, K2.6, K2.7 Code and K2.8 Preview — and K3.1 Pro is not a new base model at all, but a post-trained version of Kimi K3. This guide maps the whole family in plain language: the long-context origin story, every generation, what is open and what is closed, what each model costs to use, and which one you should actually pick.
Two corrections belong at the top, because most explainers written even a few weeks ago get them wrong. First, Kimi’s famous long-context lead has closed. The K3 series offers a 1,048,576-token context window — one million tokens — but so do the current Claude and Gemini flagships. Kimi was first to a million tokens, and it still offers the window openly and cheaply, but the window alone is no longer the difference. Second, there is no independent test that pits Kimi against Claude and Gemini on the same long documents. This guide describes the capability and its limits honestly rather than inventing test results. Prices in this guide are date-stamped “checked 18 November 2026”, because Moonshot can change them, and the pricing page is the authority if the figures here have moved.

What Kimi Actually Is — One Name, Three Different Things
<img src="

” alt=”One name on three different doors, a chat bubble, a phone app and a server rack, one brand, three things. Symbols only, no words.”>
“Kimi” is three things at once, and the first step to understanding it is knowing which one is meant. The first meaning is the assistant: the chatbot you talk to at kimi.com and in the Kimi app on iOS, Android and HarmonyOS. You ask it questions, upload documents, and it answers, summarises, researches and writes. The second meaning is the product line around that assistant: Kimi Work, a desktop app for document-centred work; Kimi Code, a terminal coding agent; Kimi Business and Enterprise accounts; and features such as Deep Research, Kimi Slides, Agent Swarm and, new with the latest launch, 3D Make. The third meaning is the model family itself — the K-series models that power all of it, from the original Kimi chat model of 2023 through k1.5, the K2 series, K3 and now K3.1 Pro.
None of these meanings is wrong. The trouble starts when a guide uses one meaning and the reader hears another — when “Kimi is open source” is taken to mean the assistant is downloadable, for example, when it is the model weights that are open and the flagship that is not fully open in the way the phrase suggests. Or when “Kimi has a million-token context” is taken to mean the free chat app gives you a million tokens, when the full window belongs to the newest models and the higher membership tiers. Keep the three meanings apart and the rest of the family tree falls into place.
This guide is the long, current version of the story. Moonshot’s assistant has been covered here before in shorter form: our short explainer “What Is Kimi? Moonshot AI’s Assistant And The K-Series Models” introduced the assistant and the long-context breakthrough in 705 words, and our short K3 explainer “Kimi K3 Explained: The Open-Weight Giant From Moonshot” covered the K3 launch. Both predate K3.1 Pro and the K2 retirement, and this guide does not repeat them at length. Its new material is the K3.1 Pro launch and its open weights, the K2 retirement and what replaces it, the pricing and access routes, the licence map, and the honest comparison with Claude and Gemini. The short “what is Kimi” section above is deliberately brief for exactly that reason — the news is elsewhere in this guide.
The Company Behind Kimi: Moonshot AI in Two Minutes

Moonshot AI was founded in early 2023 in Beijing — March 2023 by most accounts, with the operating company’s Chinese registry giving April 2023 — by Yang Zhilin, Zhou Xinyu and Wu Yuxin, Tsinghua alumni. Yang, the chief executive, holds a bachelor’s degree from Tsinghua and a PhD from Carnegie Mellon University, and he was the first author on two papers that shaped modern AI: Transformer-XL and XLNet, both about handling long sequences of text. Chinese business press also names Zhang Yutao as co-founder and chief technology officer. The company is backed by Alibaba, which led a round of about US$1 billion in February 2024, along with Tencent and Sequoia China, as reported.
That founding story explains the product. A lab led by a researcher whose best-known work was about long context built a company around long context. Moonshot’s Chinese name, 月之暗面 — “The Dark Side of the Moon” — is a statement of ambition from a small Beijing startup that set out to build a frontier model company from scratch. The infrastructure behind the product is a serving system called Mooncake, which separates prompt processing from token generation and treats the model’s memory of the conversation — its KV cache — as a distributed resource. That is the plumbing that makes very long conversations practical rather than merely possible.
Moonshot’s bet was that the constraint that mattered most was not how clever a model was on a short question, but how much it could hold in its head at once. In 2023, when the leading chatbots worked with windows of 8,000 or 32,000 tokens, a model that could read a whole book was a different kind of tool — and the market agreed, which is the story of the next section.
The Long-Context Origin Story: 200,000 Characters to 2 Million

Kimi Chat launched on 9 October 2023 with a context window of about 200,000 Chinese characters — roughly 128,000 tokens. At the time, ChatGPT’s window was 8,000 or 32,000 tokens, and the leading Chinese models, ERNIE and Qwen, were generally at 8,000. Kimi could read roughly four to sixteen times as much as its rivals in a single conversation. It opened more broadly in November 2023 after a period of invitation-only testing.
Then, on 18 March 2024, Moonshot expanded the window to 2 million Chinese characters — invited beta, on the web and in the app. That was the moment Kimi became a phenomenon in China. A chatbot that could swallow a whole novel, a whole codebase, or a shelf of research papers in one conversation was not a slightly better chatbot; it was a different instrument. By mid-2024, about 10 million monthly active users were reported — a reported figure, not a verified one, and stated here as reported only.
Why did long context matter so much? Because most real work is long. A lawyer’s case file is not one page. A researcher’s literature review is not one paper. A programmer’s project is not one function. Before long-context models, the workaround was retrieval: chop the documents into pieces, search for the relevant ones, and feed only those to the model. Retrieval works, but it loses things — the connection between two distant passages, the structure of the whole document, the detail that does not match any search query. A model that can read the whole thing does not need to guess which pieces matter. That was Kimi’s pitch in 2023 and 2024, and it worked.
The pitch has a second half, and it matters just as much: a long window is only useful if the model can actually use it. Reading a million tokens is not the same as remembering a million tokens, and every model — Kimi included — loses recall as the context grows. The honest framing, which this guide keeps throughout, is that long context is a capability with limits, not a magic number. The section on what a million tokens actually lets you do takes up those limits in full.
The K-Series Family Tree at a Glance

The K-series has moved through ten generations in three years. The table below is the whole story in one place: generation, date, the headline, and whether you can download it or only call it through an API. The sections that follow take each generation in turn.
| Generation | Date | Headline | Open or API-only |
|---|---|---|---|
| Kimi (chat) | Oct 2023 | The original assistant; 200,000 Chinese characters, then 2 million | Hosted app |
| k1.5 | 20 Jan 2025 | Multimodal reasoning model, trained with large-scale reinforcement learning | API-only, proprietary |
| Kimi K2 | 11 Jul 2025 | 1T-parameter agentic model; the line that made Kimi a model company | Open weights, Modified MIT |
| Kimi K2 Thinking | 6 Nov 2025 | Reasoning variant that interleaves thought with tool calls | Open weights |
| Kimi K2.5 | 27 Jan 2026 | First multimodal K2; vision via the MoonViT encoder | Open weights; hosted API discontinued 31 Aug 2026 |
| Kimi K2.6 | 20 Apr 2026 | General-model refresh of K2.5; thinking and non-thinking modes | Open weights; retiring end Dec 2026 |
| Kimi K2.7 Code | 12 Jun 2026 | Coding specialist built on K2.6; thinking always on | Open weights; retiring end Dec 2026 |
| Kimi K2.8 Preview | 11 Sep 2026 | Coding preview inside Kimi Code only; up to 1M context | Proprietary, Kimi Code only |
| Kimi K3 | 16 Jul 2026 | 2.8T-parameter flagship; 1M context; the base for K3.1 Pro | Open weights, Kimi K3 License |
| Kimi K3.1 Pro | 23 Oct 2026 | Post-trained K3; current flagship; open weights 4 Nov 2026 | Open weights, custom K3.1-Pro License |
Read the table with two warnings. First, the version numbers are not a quality ranking. K2.5, K2.6 and K2.7 are sub-generations of the K2 era, and the jump from one to the next is smaller than the jump from K2 to K3. A bigger number does not automatically mean a better model for your job. Second, the open/API-only column is the column that matters most. In the K2 line, the hosted API is being retired while the open weights stay downloadable — so “retired” and “gone” are not the same thing, a distinction the retirement section makes precise.
The timeline also explains why older Kimi models still appear in so many tutorials. K2 was widely deployed, and its smaller size runs on hardware that K3 cannot touch, so it remains in use long after its successor arrives. Nothing in this guide treats an older generation as obsolete on the day its successor ships. The question is always what the model is for, and the verdict section at the end answers it by use case rather than by date.
Kimi K3.1 Pro: The New Flagship, Launched 23 October 2026

Kimi K3.1 Pro launched on 23 October 2026, and it is the current flagship of the Kimi family. Moonshot’s own words, from its tech blog “Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence”, which the researcher read in full on 18 November 2026: it is “the first model in our new K3.1 series and our most capable 3T-class open-weight model to date, with a 1-million-token context window.” A faster twin, Kimi K3.1 Pro Highspeed, delivers two to three times the speed, according to Moonshot’s claim.
The most important fact about K3.1 Pro is what it is not. It is not a new base model. It is a post-trained version of Kimi K3, “with substantially larger post-training compute” — large-scale joint reinforcement learning “across hundreds of downstream domains”. The base model underneath is K3, launched in July 2026 and covered in the next section. K3.1 Pro is K3 after a much bigger round of training on the tasks people actually ask it to do.
At launch on 23 October, K3.1 Pro was available on kimi.com and in the Kimi mobile apps, in Kimi Work (desktop app version 3.1.0 and above), in Kimi Code (select it with the /model command), and through the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Enterprise access through Kimi Business is off by default — administrators must enable it. Zero Data Retention is available for API customers. And a new feature arrived with the launch: 3D Make, on kimi.com and in the Kimi app, which previews and edits 3D models, downloads them for 3D printing, and exports them as video or Live Photos.
The open weights came later, and the two dates must not be conflated. The weights were published on 4 November 2026 on Hugging Face, in the repository moonshotai/Kimi-K3.1-Pro, ungated: about 1.56 terabytes, 96 safetensor shards, MXFP4 weights produced by quantisation-aware training. Moonshot announced the release on X at 16:00 UTC on 4 November. Our news post on the release gives the day-one story; this section is the durable reference. Third-party hosting followed quickly: Amazon Bedrock added Kimi K3.1 Pro (Highspeed) on 6 November 2026, with cross-Region inference in all Bedrock regions, and Microsoft Foundry also lists it.
The architecture figures come from the Hugging Face model card, via secondary reads — Moonshot’s launch blog itself does not state the parameter counts, so they are attributed here to the model card, not to Moonshot. K3.1 Pro is a natively multimodal mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token, 896 routed experts (16 plus 2 shared per token), a 401-million-parameter MoonViT-V2 vision encoder, a 1,048,576-token context window, and Kimi Delta Attention combined with Gated DeepSeek Sparse Attention. The context figure is the one that matters most to a reader: one million tokens, in every K3-series model.
The price, as listed on Moonshot’s API pricing page and checked 18 November 2026, is $2.00 per million input tokens and $10.00 per million output tokens, with cached input at $0.20. The Highspeed twin costs $4.00 in and $20.00 out. K3.1 Pro is cheaper than its predecessor K3, which costs $3.00 in and $15.00 out, with cached input at $0.30. Batch API inference for K3.1 Pro and K3 is billed at 60% of the standard price, according to Moonshot’s pricing documents. API access requires a successful top-up of at least US$1.
On independent standing, Artificial Analysis measured an Intelligence Index of 55 for K3.1 Pro at launch, placing it fourth overall — behind Claude Opus 5.5 at 58, Claude Sonnet 5.5 at 56 and Muse Wonder Full at 56 — according to Artificial Analysis via secondary reads. Artificial Analysis later measured 67% on Terminal-Bench 4.0 and 73% on AutomationBench-AA, as reported on 4 November 2026. These are independent measurements reported by secondary sources, not verified on Artificial Analysis itself, and they are stated as such.
Moonshot’s own benchmark figures, from the tech blog read in full, are vendor claims and must be presented as vendor claims, never as independent verdicts. Terminal-Bench 4.0: 65.7%, against Claude Opus 5.5 at 66.4% and GPT-6 Astra at 58.2%, using the Claude-Code harness at maximum reasoning effort. Kimi Code Bench 3.0 Long-horizon: 61.2%, against 55.95% for K3. OfficeQA Pro v2: 64.4, which Moonshot calls state-of-the-art. BenchCAD: 92.7% with tools, against Opus 5.5 at 96.2% and GPT-6 Astra at 95.9%. On its own Kimi Hallu Bench Hard, the claim-level hallucination rate is 8.3%, against 23.4% for K3 — with the blog’s own caveat that this is a stress test and “the reported rates do not represent overall error rates in daily use”. On internal finance benchmarks, Financial Modeling improved by 7.2 points over K3, Financial Research by 11.6 points, and Business Operation by about 4.2 times. No benchmark league table in this guide is presented as a verdict, and no head-to-head verdict against Claude, GPT or Gemini is stated as a fact.
The most-discussed section of the launch blog is the cybersecurity one, and it deserves a careful reading. As post-training scaled up, Moonshot found cyber-related content in the mid-training corpus and cyber tasks in the reinforcement-learning mix, so vulnerability-research scores rose. For the open-weight release, Moonshot identified and removed cyber-related data from both training stages while keeping general coding tasks. On its internal xdev-bench-hard test, the filtered model scores 81.0% pass@3, against 90.5% for the same pipeline without the removal — a deliberate capability trade-off. Moonshot states a principle: “not to release open weights that would extend the absolute frontier of cyber capability.” That is Moonshot’s stated rationale, reported here as Moonshot’s stated rationale, not as an independent safety verdict.
Kimi K3: The 2.8-Trillion-Parameter Base It Was Built On

Kimi K3 launched on 16 July 2026, and it is the base on which K3.1 Pro was built. Its open weights were released on 26–27 July 2026 — Moonshot had promised 27 July, and multiple sources record the weights landing on 26 July — as an approximately 1.56-terabyte MXFP4 checkpoint on Hugging Face, in the repository moonshotai/Kimi-K3, under the Kimi K3 License. At launch it was described as the largest open-weight model released to date. Our short explainer “Kimi K3 Explained: The Open-Weight Giant From Moonshot” covered the launch in 722 words; this section is the durable reference, not a repeat of it.
The specifications, as reported consistently by multiple secondary reads of the launch, are: 2.8 trillion total parameters with 104 billion active per token, in a sparse mixture-of-experts design; a 1,048,576-token context window; native text, image and video input; always-on reasoning, with effort levels of low, high and max; and Kimi Delta Attention, a hybrid linear-attention design that is Moonshot’s successor to the linear-attention work from Kimi-Linear, a 48-billion-parameter model from late June 2026. The researcher did not open Moonshot’s K3 launch post first-hand, so K3 facts in this guide rest on consistent secondary reads, and no K3 launch-blog URL is cited.
At launch, K3 scored 57.1 on the Artificial Analysis Intelligence Index version 4.1, as vendor- and third-party-reported. On version 4.3, dated 7 September 2026, K3 at maximum reasoning effort scored 44. Index versions are not comparable, and this guide does not mix the two scores. The API price, as listed on Moonshot’s pricing page via secondary reads, is $3.00 per million input tokens and $15.00 per million output tokens, with cached input at $0.30 — checked 18 November 2026.
In the Kimi app, K3 runs at a 256,000-token context from the Plus membership at $19 a month, with the full 1-million-token window from the Pro membership at $39 a month and above. The membership tiers are covered in full in the access section below. K3 still matters for two reasons. First, it is the open-weight base: anyone who wants to download, fine-tune or self-host the current Kimi generation starts with K3 or with its post-trained successor, K3.1 Pro. Second, it is the cheaper of the two to run only in the sense that it costs more per token than K3.1 Pro — the successor is both newer and cheaper, which is unusual, and it is the reason K3.1 Pro, not K3, is the default recommendation in the verdict section.
The self-hosting reality belongs here too. K3’s weights are about 1.5 terabytes — roughly 1,509 gigabytes at 4-bit precision, a third-party estimate stated here as approximate. No single graphics card or standalone machine can run it. It is a data-centre and cloud model, not a laptop download, and anyone planning to self-host should budget for that before downloading anything.
The K2 Era: How Kimi Became an Agentic Model Company

Kimi K2 launched on 11 July 2025, and it is the model that turned Moonshot from a long-context lab into an agentic model company. Its headline figures, from the technical report arXiv:2507.20534 as read via consistent secondary sources, are: 1 trillion total parameters with 32 billion active per token, in a mixture-of-experts design with 384 experts, 8 routed plus 1 shared per token; 15.5 trillion pre-training tokens; a 128,000-token context at launch, expanded to 256,000 tokens with the Kimi-K2-Instruct-0905 update in September 2025; and a Modified MIT licence, with open weights on Hugging Face. It was trained with the MuonClip optimiser, with reportedly zero loss spikes during training. At launch it was the highest-ranked open model on the LMSYS Chatbot Arena and fifth overall, as vendor- and third-party-reported.
What made K2 different was not its size but its job. It was built for agentic work: multi-step tool use, autonomous task execution, the kind of work where a model plans a task, calls tools, reads the results, and carries on without a human steering every step. Long context was the enabler — an agent that forgets the plan halfway through is not an agent — but the agentic training was the point. K2 was the model that made “Kimi” a name in the developer world, not just in the Chinese consumer market.
Kimi K2 Thinking followed on 6 November 2025: a reasoning variant that interleaves chain-of-thought with tool calls, with a 262,144-token context and native INT4 quantisation, reported able to run 200 to 300 consecutive tool calls without drifting. Its open weights were published. The API price, via OpenRouter and third-party trackers checked 17 November 2026, is $0.60 per million input tokens and $2.50 per million output tokens. A faster variant, kimi-k2-thinking-turbo, also exists, at $0.60 in and $2.40 out according to third-party trackers — a secondary figure, flagged here as reported.
Both models are now history in the hosted sense, and the next sections explain why. But their influence survives: the agentic training, the tool-use patterns and the long-horizon coding work that K2 pioneered are the direct ancestors of what K3.1 Pro does today. For readers comparing coding agents more broadly, our comparison of AI coding assistants sets out how the Kimi coders sit against the rest of the field.
K2.5, K2.6 and K2.7 Code: Vision, Longer Coding, and the Coding Specialist

The K2 line kept moving through the first half of 2026, in three steps that are best understood as one compact story: vision, then a general refresh, then a coding specialist.
| Model | Date | What changed | Context |
|---|---|---|---|
| Kimi K2.5 | 27 Jan 2026 | First multimodal K2; vision via a ~400M-parameter MoonViT encoder; continued pre-training on ~15T mixed visual and text tokens | 262,144 tokens |
| Kimi K2.6 | 20 Apr 2026 | General-model refresh of K2.5; same 1T/32B shape; native vision; thinking and non-thinking modes | 262,144 tokens |
| Kimi K2.7 Code | 12 Jun 2026 | Coding specialist built on K2.6; thinking always on | 262,144 tokens |
Kimi K2.5, launched 27 January 2026, was the first multimodal K2: text, image and experimentally video input, through a roughly 400-million-parameter MoonViT vision encoder, with continued pre-training on about 15 trillion mixed visual and text tokens from the K2 base. The Agent Swarm preview followed on 9 February 2026. Its open weights were published. On Moonshot’s own platform, K2.5 was discontinued on 31 August 2026, together with the original moonshot-v1 model; Azure keeps it until 26 January 2027.
Kimi K2.6, launched 20 April 2026, was a general-model refresh of K2.5 in the same 1-trillion-total, 32-billion-active shape, with native vision and thinking and non-thinking modes, and improved long-horizon coding and agent work. Its open weights were published. The API price, as listed on Moonshot’s pricing page via secondary reads, is $0.95 per million input tokens and $4.00 per million output tokens, with cached input at $0.16 — checked 18 November 2026.
Kimi K2.7 Code, launched 12 June 2026, was the coding specialist built on K2.6, with thinking always on. Its API identifiers are kimi-k2.7-code and kimi-k2.7-code-highspeed, and its open weights were published under the Modified MIT licence. The API price is $0.95 in and $4.00 out per million tokens, with cached input at $0.19.
One more K2-generation model deserves a line, and only a line. Kimi K2.8 Preview rolled out on 11 September 2026 inside Kimi Code only, behind the existing kimi-for-coding model identifier, with up to a 1-million-token context and low, high and max reasoning effort. It is proprietary — not open-weight — and it must not be presented as a downloadable model. Beyond the rollout date and the context figure, its benchmark claims were not verified in the sources read for this guide, so nothing more is claimed about it here.
The K2 Retirement: What Ends in December 2026 — and What Replaces It

The K2 story has an ending, and it arrived fast. Moonshot’s notice on its API Model List and pricing page, quoted identically by several secondary sources that read the official pages, says: “Starting October 31, 2026, K2 series models will no longer be available to newly registered users; existing users are not affected. We plan to retire all K2 series models by the end of December 2026.” The named migration target is kimi-k3.1-pro. The researcher did not open the Model List page first-hand, so the Director is asked to spot-check the notice on platform.moonshot.ai before sign-off.
The practical effect is simple. A new API account created after 31 October 2026 can only use the K3 series. Existing accounts keep K2.6 and K2.7 Code until the planned end-of-December retirement — and there is no exact shutdown day published, because this is a plan, not a timestamp, and this guide names no date. The earlier deprecations set the pattern: the original kimi-k2 series was discontinued on the hosted API on 25 May 2026, kimi-latest on 28 January 2026, and kimi-k2.5 together with moonshot-v1 on 31 August 2026.
Two things survive the retirement. First, the open weights stay available after the hosted retirement. A K2 model you downloaded does not stop working because Moonshot stops serving it — the weights are yours, under the Modified MIT licence, and they keep running wherever you run them. Second, the ideas survive: the agentic training and tool-use patterns of the K2 era are the foundation of K3.1 Pro.
But the migration is not a drop-in swap, and this guide says so plainly. K3.1 Pro’s thinking is always on, its reasoning effort defaults to maximum, and its sampling settings are fixed, according to reads of Moonshot’s documentation. A system built around K2’s switchable thinking modes and adjustable sampling will behave differently on K3.1 Pro — possibly better, possibly not, but differently. Migration needs testing, not a one-line model-identifier change. Anyone with a K2-based production system should test K3.1 Pro against their own tasks before the end of December, while K2 is still available to compare against.
What 1 Million Tokens Actually Lets You Do — and What It Doesn’t

One million tokens is a number that needs translating. As a rough rule of thumb, a token is about three-quarters of an English word, so a million tokens is roughly 750,000 words — presented here as approximate, not exact. That is about ten average novels, or a whole medium-sized codebase, or a shelf of research papers, or a long legal or financial document set, in a single conversation. In the K3 series, that window is real and current: Moonshot’s own pricing documents state that “All K3 series models support a 1,048,576-token context window.” For background on the retrieval systems that long context partly replaces, our explainer of BERT, Llama, Whisper and RAG covers the embedding and retrieval ideas behind the alternative approach.
What does that let you do? Feed a whole codebase to Kimi Code and ask it to refactor across files, with the model able to see the imports, the tests and the call sites at once. Hand a lawyer’s full case file to the assistant and ask for the contradictions between documents. Give a researcher a hundred papers and ask for the synthesis, with the model able to check a claim against the source paper rather than against its memory. Give a finance analyst a year’s worth of filings and ask for the trend. These are the jobs Kimi was built for, and they are the jobs where a long window changes what is possible, not just what is convenient.
What it does not do is just as important. First, long-context recall degrades with length on every model. A model that reads a million tokens does not remember all million equally well; the middle of a very long context is where details go missing, and no Kimi-specific measured figure for that degradation was found in the sources read for this guide — so none is stated here. Second, a million tokens costs a million tokens: at K3.1 Pro’s price of $2.00 per million input tokens, a full window costs about $2.00 to fill, and the output costs $10.00 per million tokens on top. Long context is cheap per token, but a million tokens is still a million tokens. Third, for the very largest corpora — a company’s entire document store, a decade of archives — retrieval is still needed. A million tokens is enormous for a conversation and small for a library.
And one thing this guide deliberately does not do: it does not report a test. No independent head-to-head long-context test of Kimi against Claude and Gemini on the same documents was found as of 18 November 2026 — no LongBench, RULER or BABILong-style comparison with all three on the same documents. No test results, accuracy figures or “tested on a 500-page PDF” claim appears anywhere in this guide, because none exists to report. The full tested comparison belongs to a separate queue item, “Long-Context AI Compared”, and is out of scope here.
Kimi vs Claude vs Gemini on Long Documents: The Honest 2026 Picture

The comparison most readers want is with Claude and Gemini, and the honest 2026 answer starts with a correction. Kimi’s 1-million-token window is real and current. But the lead has closed. According to the sibling evergreen briefs in this same programme, verified on 13 November 2026, current Claude models — Opus 5.5 and Fable 5.2 — offer a 1-million-token context with 128,000 tokens of maximum output, and Gemini 4 Carbon offers 1 million tokens of input context. The old comparison, Kimi at 1 million against Claude at 200,000, is out of date: Claude’s 200,000-token figure belongs to older models such as Haiku 4.5. This guide does not write “Kimi has five times Claude’s context”, because it is no longer true.
So where does Kimi still differ? On three things. First, open weights. Kimi K3.1 Pro’s weights are downloadable — about 1.56 terabytes of them — while Claude and Gemini are closed models you can only rent through an API. For a team that needs to run a model on its own hardware, audit it, fine-tune it, or keep its data off a vendor’s servers, that is a real difference, not a marketing one. Second, price. At $2.00 per million input tokens and $10.00 per million output tokens, K3.1 Pro is a fraction of the cost of the closed frontier models, and it is cheaper than its own predecessor K3. Third, agentic coding. The K2 era made Moonshot an agentic model company, and K3.1 Pro’s strongest vendor-reported results — Terminal-Bench 4.0 at 65.7%, Kimi Code Bench 3.0 Long-horizon at 61.2% — are in coding and agent work, where it sits close to Claude Opus 5.5’s 66.4% on Terminal-Bench. Those are Moonshot’s own figures, stated as Moonshot’s own figures, not an independent verdict.
The wider field matters too. Kimi is not the only open-weight family. DeepSeek, its Chinese peer, is covered in our explainer of DeepSeek, and Qwen, Alibaba’s open-weight family, in our explainer of the Qwen models — Qwen’s newest flagship, Qwen4-Max, is API-only, so its open line sits a generation behind, much like Kimi’s did before K3. For the Claude and Gemini sides of the comparison, the sibling explainers are the right next reading: our explainer of the Claude model family and our explainer of the Gemini model family. And for the decision itself — which model to choose for a given job, and what speed, cost, reasoning depth and context actually trade off against each other — our guide to choosing an AI model sets out the framework.
The short version is that Kimi was first to a million tokens and still offers it openly and cheaply; the frontier closed models have caught up on window size, so Kimi’s remaining edges are open weights, price and agentic coding — not the window alone. Any guide that declares a winner between Kimi, Claude and Gemini on the basis of benchmark scores has not earned the verdict, because no independent head-to-head result was verified first-hand during this research.
Open or Closed? Licences and Access, Decoded

“Is Kimi open source?” is the most-asked question about the family, and the honest answer has three parts. First, most Kimi weights are open to download. Second, the licences have exceptions, and they sit exactly where the money is. Third, open weights are not the same as open source — and the difference matters for anyone building a business on Kimi.
The K2 line — K2, K2 Thinking, K2.5, K2.6 and K2.7 Code — is released under the Modified MIT licence, which is permissive: the weights are downloadable and commercial use is allowed. But the training data and the full training code were not released. K3 is released under the Kimi K3 License, a custom licence in the modified-MIT family. K3.1 Pro is released under a custom Kimi K3.1-Pro License — not a standard permissive licence. As published on the Hugging Face model card, it grants use, copying, modification, deployment, fine-tuning, distribution and selling, with conditions that apply only to very large operators: model-as-a-service businesses with more than US$20 million in revenue over 12 months need a separate agreement for commercial use, and products with more than 100 million monthly users or more than US$20 million in monthly revenue must display attribution. Those thresholds are stated here as published on the Hugging Face model card, and the card itself is linked in the Sources section; this guide does not paraphrase the licence’s legal effect. Two Kimi models are not open at all: K2.8 Preview is proprietary, with no open weights, and k1.5 is proprietary, with no weights published.
The practical rule is simple. Read the licence file on the model’s Hugging Face card before any commercial use, confirm whether the model you want is open or hosted, and never assume that what is true of K2 is true of K3.1 Pro. The licence has tightened as the models have grown more capable, and the direction of travel is toward more conditions at the top of the range, not fewer.
Then there is the phrase “open-weight” itself. Open weights mean you can download the trained parameters and run the model. They do not mean the training data is published, the training code is published, or the model is open source in the sense the Open Source Initiative defines it. Kimi is open-weight, not open source, and the distinction matters commercially: you can build on the weights, but you cannot audit how they were made.
How to Try Every Kimi Model: App, API, or Your Own Machine

There are three ways into Kimi — the app, the API, or your own machine — and the right one depends on what you want to do, not on which is most impressive. Start with the easiest, and move down the list only when the easier route stops being enough.
The app route is kimi.com and the Kimi app on iOS, Android and HarmonyOS. A free tier exists, and no credit card is required for casual use, according to secondary reads of the help centre — but the free tier’s limits are not published, and this guide states no number of free messages or runs, because none was verified. The paid memberships, in the international US-dollar list as read in September and October 2026, are Plus at $19 a month ($15 on the annual plan), Pro at $39 ($31), Max at $99 ($79) and Ultra at $199 ($159). These names replaced the old tempo names — Moderato at $19, Allegretto at $39, Allegro at $99 and Vivace at $199 — around 28 September 2026, with prices unchanged, according to secondary sources; the rename was not verified on the live kimi.com pricing page in this session, so the membership names and prices here are as reported, and the live page is the authority. Never mix the old and new names in one list, and never mix the US-dollar list with the separate China pricing in yuan. K3 runs at a 256,000-token context from Plus, with the full 1-million-token window from Pro and above. Kimi Code access is tied to membership, and its credits multiplier rises with the tier. The app route is the right first step for anyone who wants to know what Kimi feels like: ask it questions, upload a document, try Deep Research, and see whether the answers are useful.
The API route is the Kimi API Platform — platform.moonshot.ai internationally, and platform.kimi.com and moonshot.cn in China — with the base address api.moonshot.ai internationally, in an OpenAI-compatible format. It is pay-as-you-go and prepaid, and it is separate from membership: a paid membership does not discount API tokens. Access requires a successful top-up of at least US$1. File upload and extraction APIs are temporarily free, according to Moonshot’s pricing documents read on 18 November 2026. Context caching is automatic, with a default five-minute time-to-live, and cache writes are billed separately by time-to-live — five minutes or one hour. The price table, as listed on Moonshot’s API pricing page via consistent secondary reads and checked 18 November 2026, is: Kimi K3.1 Pro at $2.00 in and $10.00 out per million tokens, cached input $0.20; Kimi K3.1 Pro Highspeed at $4.00 in and $20.00 out; Kimi K3 at $3.00 in and $15.00 out, cached input $0.30; Kimi K2.6 at $0.95 in and $4.00 out, cached input $0.16; Kimi K2.7 Code at $0.95 in and $4.00 out, cached input $0.19; and Kimi K2 Thinking at $0.60 in and $2.50 out, cached input $0.15, via third-party trackers. Batch API inference for K3.1 Pro and K3 is billed at 60% of the standard price. For the wider question of how API access compares across providers, our explainer of the OpenAI API covers the patterns — keys, pricing, rate limits — that apply to Moonshot’s platform as much as to any other provider. Third-party hosts also serve Kimi models: OpenRouter, Amazon Bedrock, which added K3.1 Pro on 6 November 2026, and Microsoft Foundry, which also lists it.
The local route is the open weights, from Hugging Face in the moonshotai organisation and from GitHub, run through vLLM, SGLang, TensorRT-LLM or llama.cpp — K3 text support was merged into llama.cpp on 15 August 2026, and GGUF builds exist via Unsloth. No GGUF release for K3.1 Pro was found as of 18 November 2026, and this guide does not claim one exists. The realistic self-hosting picture is a data-centre one: about 1.5 terabytes of weights for K3 or K3.1 Pro, at roughly 1,509 gigabytes in 4-bit precision — a third-party estimate, stated as approximate. Start small, confirm the model runs, and only then move up in size. The desktop route sits between the app and the API: Kimi Work, the desktop app for Windows and Apple silicon Macs, version 3.1.0 and above for K3.1 Pro, is a document-centred workspace with Deep Research, Slides and Agent Swarm, while Kimi Code is the open-source terminal coding agent, with K3 and K3.1 Pro selectable via the /model command.
The order of operations for a new user is: try kimi.com first, move to the API when you need to build with it, run weights locally when you need control or privacy, and install Kimi Code when the job is writing software rather than chatting about it. One caveat belongs with the app route: this guide does not state that Kimi is available worldwide, and it names no list of countries, because availability by country and sign-up requirements — whether a phone number is needed, which regions are blocked, whether the free tier is available in every region — were not verified in this session. The honest phrasing is that Kimi is available in English at kimi.com and through the international API platform, and that regional differences exist.
Which Kimi Should You Pick? A Verdict by Use Case — and What to Watch Next

There is no single best Kimi. There is a best Kimi for each job, and the verdict below is by use case — the only honest way to end a guide about a family of ten generations.
| Use case | Pick | Why |
|---|---|---|
| Casual chat | kimi.com free tier | Free to try, latest models, no setup |
| Long-document work | Kimi K3.1 Pro in the app (Pro tier and above for the full 1M window) | Current flagship; full million-token context from Pro |
| Coding, local or agentic | Kimi Code with K3.1 Pro | Current coding flagship; K2.7 Code is retiring |
| API production | kimi-k3.1-pro | Newest, and cheaper per token than K3 |
| Running a model locally | Kimi K3 or K3.1 Pro weights — data-centre hardware only | Current open generation; no laptop-scale option |
| Small local models | None in the Kimi family | Kimi has no small downloadable model; look elsewhere |
| Enterprise | Kimi Business, with K3.1 Pro enabled by an administrator | Team accounts; K3.1 Pro is off by default |
The limits belong in the same breath as the recommendations. K3.1 Pro is less than a month old as this guide is checked, and its benchmark evidence is Moonshot’s own or secondary-reported; treat early claims about it with the scepticism any one-month-old model deserves. The K2 series is retiring at the end of December 2026, so a system built on K2 needs a tested migration, not a one-line change. The licences have conditions at the top of the range, so the licence file must be checked per model before commercial use. And the membership plan names changed in September 2026, so an older guide’s tempo names — Moderato, Allegretto, Allegro, Vivace — describe the same prices under different labels.
What to watch next is short, and every item on it is unconfirmed. The successors in the K3.1 series, which Moonshot’s blog implies but has not named — the moment one appears, the flagship moves again. The K2 shutdown at the end of December 2026, with no exact day published. And Moonshot’s push beyond China, including its expansion into Singapore in 2026, which may change the availability picture this guide deliberately leaves open. None of these has a date. The family moves every few months, and the check that matters is the one this guide performed on 18 November 2026: read the release notes, read the licence file, and do not trust a guide — including this one — past its checked date. Checked 18 November 2026.
Sources
All sources below were read on 18 November 2026, except where noted. Facts, dates, prices and licence details in this guide come from these pages. No competitor or comparison-site pages were used as sources. Prices checked 18 November 2026.
- Moonshot AI — “Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence” (launch 23 Oct 2026; benchmarks; cyber-data removal; availability; 3D Make; blog page updated 24 Oct 2026) — read in full, all 282 lines — https://www.kimi.com/en/blog/kimi-k3-1-pro
- Kimi API Model Inference Pricing — context windows 1,048,576 / 262,144, batch inference at 60%, caching terms, file APIs temporarily free; updated 12 Nov 2026 — read; the price table itself is JavaScript-rendered and did not appear in the text fetch, so prices are as listed via consistent secondary reads — https://platform.moonshot.ai/docs/pricing/chat
- Hugging Face — Kimi-K3.1-Pro model card and licence (4 Nov 2026 open weights, ~1.56 TB, 96 safetensor shards, MXFP4, licence thresholds — read via secondary reads, cross-checked: accessallgpt.com, dev.to, theindependentai.com — consistent; licence text not opened first-hand) — https://huggingface.co/moonshotai/Kimi-K3.1-Pro
- Hugging Face — Kimi-K3 model card (26–27 Jul 2026 weights, ~1.56 TB MXFP4 checkpoint, Kimi K3 License — read via secondary reads: aistackcurrent.com, theairankings.com, wan27.org, ainchina.com — consistent on 16 Jul 2026 date, 2.8T/104B, 1M context) — https://huggingface.co/moonshotai/Kimi-K3
- Moonshot API Model List / pricing page — K2 retirement notice (“Starting October 31, 2026, K2 series models will no longer be available to newly registered users… retire all K2 series models by the end of December 2026”) — as quoted by dev.to, bettertoken.ai, aiproleague.com, ainewsway.com, benchlm.ai, maximem.ai — identical wording; Model List page not opened first-hand
- Kimi K2 technical report (arXiv:2507.20534; 11 Jul 2025, 1T/32B, 15.5T pre-training tokens, Modified MIT, MuonClip) — read via secondary reads: aiwiki.ai, intuitionlabs.ai, leanware.co — consistent
- Kimi k1.5 technical report (arXiv:2501.12599; 20 Jan 2025, 128K context, large-scale reinforcement learning) — read via secondary reads: ai-tldr.dev, examsnap.com
- K2 Thinking / K2.5 / K2.6 / K2.7 Code / K2.8 Preview timeline (6 Nov 2025; 27 Jan 2026; 20 Apr 2026; 12 Jun 2026; 11 Sep 2026 — read via secondary reads: daily.rclabs.in, benchgecko.ai, aitoolchain.io, scriptbyai.com, ai-tldr.dev, zentor.ai, lorphic.com — consistent dates)
- Moonshot AI history (Mar/Apr 2023 founding, 9 Oct 2023 Kimi Chat launch, 200,000 → 2 million Chinese characters, 18 Mar 2024) — via theteardown.co, zsjunai.github.io History of AI, canvasbusinessmodel.com, aiwiki.ai — consistent
- Kimi membership pricing (Plus/Pro/Max/Ultra rename ~28 Sep 2026; prices unchanged — via rationalgo.ai reading the help centre 16 Sep 2026; solcard.cc / UsagePricing capture 28 Sep 2026; novaforwork.com checked 19 Oct 2026; live kimi.com pricing page not checked in this session)
- Artificial Analysis Intelligence Index 55 at launch, 4th overall (via amargupta.tech, gptonline.ai, GER news post 1576); Terminal-Bench 4.0 67% and AutomationBench-AA 73% (via theindependentai.com, 4 Nov 2026)
- AWS Bedrock announcement, 6 Nov 2026 (Kimi K3.1 Pro Highspeed, cross-Region inference — via aiagentupdate.com, letsdatascience.com); Microsoft Foundry listing (GER news post 5033)
- GER news — “Kimi K3.1 Pro Open Weights” (4 Nov 2026) — https://geroundtable.com/kimi-k3-1-pro-open-weights/
- GER news — “Kimi K3.1 Pro on Amazon Bedrock” (6 Nov 2026) — https://geroundtable.com/kimi-k3-1-pro-amazon-bedrock/
- GER news — “Kimi K3.1 Pro on Microsoft Foundry” — https://geroundtable.com/kimi-k3-1-pro-microsoft-foundry/
- GER news — “Kimi K3.1 Pro Fourth on Artificial Analysis Index” — https://geroundtable.com/kimi-k31-pro-fourth-artificial-analysis-index/
- GER news — “Moonshot Launches Kimi K3.1 Pro Open Weights November 4” — https://geroundtable.com/moonshot-launches-kimi-k31-pro-open-weights-november-4/
- GER — “What Is Kimi? Moonshot AI’s Assistant And The K-Series Models” (post 507, 18 Oct 2026) — https://geroundtable.com/what-is-kimi-moonshot-ai-assistant/
- GER — “Kimi K3 Explained: The Open-Weight Giant From Moonshot” (post 474, 18 Oct 2026) — https://geroundtable.com/kimi-k3-open-weight-giant-explained/
- Sibling evergreen briefs, same programme — C5 Claude (current Claude models: 1M context, 128K max output; verified 13 Nov 2026) and C6 Gemini (Gemini 4 Carbon: 1M input context; verified 13 Nov 2026)


0 Comments