Moonshot AI
Dossier checked 28 Sep 2026
Newest frontier release
Kimi K3.1 Pro
24 Oct 2026 · the lab’s announcement
Post-trained version of Kimi K3 with a 1M-token context, live on Kimi.com, Kimi Work, Kimi Code and the Kimi API on the day; open weights promised by 4 Nov
Flagship list price, per million tokens
$3.00 in $15.00 out
Kimi K3, cache-miss input, released 16 Jul 2026 · pricing page
Kimi K3.1 Pro (24 Oct), a post-trained K3, is newer and cheaper at $2.00 in and $10.00 out; its weights are due by 4 Nov, and this page’s weights and channels are still read for K3.
Weights
Open flagship
Full weights under a bespoke Kimi K3 licence, which is not an open-source licence: a separate agreement is needed for model-as-a-service businesses with more than $20M of revenue, and attribution above 100M monthly users. Moonshot says “open weight”, never “open source”. Source
Where you can buy the flagship
- Kimi API and Kimi apps
- Hugging Face
- Alibaba Cloud Model Studio
- Amazon Bedrock
- Microsoft Foundry (via Fireworks)
- Fireworks AI
- OpenRouter
5 releases in the year to 30 Oct 2026, 3 the year before · 9 in all since 2023
The question we ask of every lab: is most of its value in the model, which gets cheaper fastest, or in something other work is built on? Our answer: in the model itself.
Moonshot’s models are open and frontier-adjacent, which is the fastest-commoditising place to stand. The durable bet is Kimi as a consumer habit in China and the agentic, long-context niche it reached first.
What would change our readKimi loses its consumer foothold to a better-distributed rival, or its open weights fall off the frontier pace.
01 · What it runs on, and the money
Trains and serves on press report
No chips and no datacentre of its own. Bloomberg reported in July 2026 that it uses about 20,000 NVIDIA Hopper chips under a compute deal with Alibaba; Alibaba denied supplying H200s but not the deal, and Reuters could not confirm it. Confirmed: day-zero serving of Kimi K3 on Alibaba Cloud and Huawei Ascend. Source
What it depends on press reportIts reported training compute comes through Alibaba, a shareholder that also sells a competing model; Alibaba has disputed only the chip type. Source
Biggest risk press reportIn July 2026 the White House science office alleged that Moonshot distilled a US model and used restricted NVIDIA servers, and the Treasury warned that such conduct could bring sanctions. Moonshot has disputed the distillation claim. Source
- Raised press report
- More than US$5.5B by May 2026, then a further round of about US$3.5B in July 2026 Source
- Valuation press report
- About US$35B at the July 2026 round; a pre-IPO round at about US$50B is reported in progress Source
- Revenue press report
- Not disclosed. About US$1B annualised in August 2026, up from about $300M in June Source
- Structure press report
- Beijing Moonshot AI Technology, private, founded in 2023 out of Tsinghua. It has moved its holding structure onshore and confidentially filed for a Hong Kong IPO. Source
NotesThe year-ago price is Kimi K2 0905 as archived in October 2025. On that basis K3 costs five times as much on input and six times on output, the highest open-weight flagship price on this site. Headcount is omitted because reported figures range from 80 to 300.
02 · Releases
data through 30 Oct 2026
- Moonshot 9 Kimi K3.1 Pro24 Oct 2026
Each dot is a frontier release at its date. Click one to find it in the full log. Every lab on one chart
The landmarks
- 16 Jul 2026 Kimi K3 2.8T-parameter MoE with 1M context; the largest open-weight release to date, under a bespoke licence
- Nov 2025 Kimi K2 Thinking Open reasoning and agentic model
- Jul 2025 Kimi K2 Trillion-parameter MoE, open-sourced
- Nov 2023 Kimi (chatbot) Long-context assistant: a closed beta from October, open to the public in November
The full release logall 9, newest first
- 2026
- 24 Oct 2026 Kimi K3.1 Pro Frontier Post-trained version of Kimi K3 with a 1M-token context, live on Kimi.com, Kimi Work, Kimi Code and the Kimi API on the day; open weights promised by 4 Nov Source
- 16 Jul 2026 Kimi K3 Frontier 2.8T-parameter MoE with 1M context; the largest open-weight release to date, under a bespoke licence Source
- 20 Apr 2026 Kimi K2.6 Frontier Agentic coding, general availability Source
- 27 Jan 2026 Kimi K2.5 Frontier Multimodal update, on Kimi.com, the app, the API and Kimi Code from launch Source
- 2025
- Nov 2025 Kimi K2 Thinking Frontier Open reasoning and agentic model
- Sep 2025 Kimi K2 Instruct 0905 Frontier Longer context, stronger coding: a new K2 checkpoint, open weights and on the API
- Jul 2025 Kimi K2 Frontier Trillion-parameter MoE, open-sourced
- 25 Jan 2025 Kimi K1.5 Frontier Reasoning model matching o1, on the global Kimi web from 25 Jan; the API was by application Source
- 2023
- Nov 2023 Kimi (chatbot) Frontier Long-context assistant: a closed beta from October, open to the public in November
03 · When to reach for them
practical read
- Long context
- Reach for Kimi K3 when the task needs a 1M-token context (Jul 2026).
- Largest open model to host
- Use Kimi K3 for the largest open-weight model you can host, 2.8T parameters (Jul 2026).
- Lower cost
- Pick Kimi K2.6 when cost matters: $0.95 in and $4.00 out per million tokens (Apr 2026).
04 · Weights, across all ten
the flagship
05 · Essays that read Moonshot AI
from the archive
06 · About Moonshot AI
at a glance
- Founded
- 2023
- HQ
- Beijing, China
- Leadership
- Yang Zhilin (CEO)
- Backers
- Alibaba, Tencent, Meituan, National AI Industry Investment Fund
- Also builds
- Kimi app, Kimi Code, Kimi Work
- Known for
- Long-context, Kimi
Founded March 2023 by Tsinghua schoolmates Yang Zhilin and others, headquartered in Beijing. Known for the Kimi assistant, an early long-context leader, and the open-weight Kimi K2 and K3 models; Kimi K3 (Jul 2026) has 2.8T parameters and a 1M-token context.