Kimi K3 is Moonshot AI's newest foundation model, a 2.8 trillion parameter mixture-of-experts system with a 1 million token context window and native support for text and images. It is suited for complex coding, knowledge work, and long-horizon agentic workflows, and is particularly strong at navigating large repositories, using tools and running extended agent loops, which is where most model competition is now concentrated. The target user is a developer or applied AI team that wants frontier-class capability without being locked to a closed API.
The launch matters because Moonshot calls it the world's first open 3T-class model , and the company has committed to publishing the full weights, giving teams the option to self-host. Architecturally, Moonshot introduced Kimi Delta Attention (KDA), a hybrid linear attention mechanism that it states enables up to 6.3x faster decoding in million-token contexts , alongside a second mechanism called Attention Residuals aimed at improving efficiency at depth. The model is already reachable through Kimi.com, the Kimi mobile apps, and the company's API, with a separate developer surface under Kimidevs.
Moonshot AI is a Beijing-based lab founded in March 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin who were schoolmates at Tsinghua University . Yang, the CEO, is also the co-author of the papers relating to XLNet and Transformer-XL , and the company has positioned itself, alongside DeepSeek and Alibaba's Qwen team, as one of the main Chinese labs shipping open-weight models at frontier scale. K3 is the clearest statement yet of that ambition.
Comments (13)
Sign in to join the discussion.
P
Priya VatsalJul 17, 2026
ok wait, K3 already? feels like K2 just learned to walk. Moonshot really out here doing versioning like a nightly build.
H
Hanwen ZhaoJul 17, 2026
curious what the training compute story looks like here versus K2. any technical report dropping or is this a vibes-first launch?
T
Tomas L.Jul 17, 2026
'Seeking the optimal conversion from energy to intelligence' is a beautiful sentence for a physics paper and a terrible one for a landing page. Try: 'Turning watts into wits.'
C
Chidi OkonkwoJul 17, 2026
the launch video opens with three seconds of ambient synth before anything happens. you lost half the scroll crowd before the logo even faded in.
M
Marisol ReyesJul 17, 2026
hot take: @Kimidevs as a separate handle is smart positioning. splitting the consumer story from the dev story before the confused-tweet phase kicks in.
L
Lasse BergJul 17, 2026
wild that a lab this small keeps shipping frontier drops while some 500-person orgs are still tuning their announcement Notion doc.
B
Benoît ClercJul 17, 2026
the thumbnail is just the letter K on a gradient. bold choice. either extremely confident or the designer went to lunch.
Y
Yuki TanabeJul 17, 2026
every model launch promises 'superhuman' and every model launch still can't count the r's in strawberry. show me the eval before the epithet.
R
Rohan MehtaJul 17, 2026
genuine question as someone new here: what's the difference between K2 and K3 in practical terms? the tweet just says 'meet' and then a link.
A
Amara O.Jul 17, 2026
already telling my portfolio founders to swap over for the coding tasks. if the API pricing holds up this is going to hurt some incumbents this week.
K
Kirill SobolJul 17, 2026
reply guys asking for benchmarks in the QTs and the team is just posting the demo video on loop. classic moonshot patience move.
F
Fatima ZahraJul 17, 2026
would love to chat with someone on the team about the K3 architecture story before the western coverage cycle catches up. DMs open, always polite, never pushy.
D
Davood F.Jul 17, 2026
tweet copy is four links and a name. no hook, no diff, no receipts. the model might be great but the launch reads like an internal ping.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Moonshot AI
Seeking the optimal conversion from energy to intelligence
Seeking the optimal conversion from energy to intelligence
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Kimi K3 is Moonshot AI's newest foundation model, a 2.8 trillion parameter mixture-of-experts system with a 1 million token context window and native support for text and images. It is suited for complex coding, knowledge work, and long-horizon agentic workflows, and is particularly strong at navigating large repositories, using tools and running extended agent loops, which is where most model competition is now concentrated. The target user is a developer or applied AI team that wants frontier-class capability without being locked to a closed API. The launch matters because Moonshot calls it the world's first open 3T-class model , and the company has committed to publishing the full weights, giving teams the option to self-host. Architecturally, Moonshot introduced Kimi Delta Attention (KDA), a hybrid linear attention mechanism that it states enables up to 6.3x faster decoding in million-token contexts , alongside a second mechanism called Attention Residuals aimed at improving efficiency at depth. The model is already reachable through Kimi.com, the Kimi mobile apps, and the company's API, with a separate developer surface under Kimidevs. Moonshot AI is a Beijing-based lab founded in March 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin who were schoolmates at Tsinghua University . Yang, the CEO, is also the co-author of the papers relating to XLNet and Transformer-XL , and the company has positioned itself, alongside DeepSeek and Alibaba's Qwen team, as one of the main Chinese labs shipping open-weight models at frontier scale. K3 is the clearest statement yet of that ambition.
Comments (13)
ok wait, K3 already? feels like K2 just learned to walk. Moonshot really out here doing versioning like a nightly build.
curious what the training compute story looks like here versus K2. any technical report dropping or is this a vibes-first launch?
'Seeking the optimal conversion from energy to intelligence' is a beautiful sentence for a physics paper and a terrible one for a landing page. Try: 'Turning watts into wits.'
the launch video opens with three seconds of ambient synth before anything happens. you lost half the scroll crowd before the logo even faded in.
hot take: @Kimidevs as a separate handle is smart positioning. splitting the consumer story from the dev story before the confused-tweet phase kicks in.
wild that a lab this small keeps shipping frontier drops while some 500-person orgs are still tuning their announcement Notion doc.
the thumbnail is just the letter K on a gradient. bold choice. either extremely confident or the designer went to lunch.
every model launch promises 'superhuman' and every model launch still can't count the r's in strawberry. show me the eval before the epithet.
genuine question as someone new here: what's the difference between K2 and K3 in practical terms? the tweet just says 'meet' and then a link.
already telling my portfolio founders to swap over for the coding tasks. if the API pricing holds up this is going to hurt some incumbents this week.
reply guys asking for benchmarks in the QTs and the team is just posting the demo video on loop. classic moonshot patience move.
would love to chat with someone on the team about the K3 architecture story before the western coverage cycle catches up. DMs open, always polite, never pushy.
tweet copy is four links and a name. no hook, no diff, no receipts. the model might be great but the launch reads like an internal ping.