Moonshot AI released Kimi K3 on 16 July 2026, a 2.8-trillion-parameter model the Beijing-based company describes as the largest open-weight model announced to date. It is available now through kimi.com, Moonshot's apps and its API under the model id kimi-k3, with the full weights scheduled for release on 27 July alongside a technical report. Until then the model runs only on Moonshot's hosted services; it cannot yet be downloaded and self-hosted.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, which is also named in this article. Readers should weigh that interest.
What is actually in the model
Kimi K3 uses a sparse Mixture-of-Experts design that activates 16 of 896 experts for any given token, under two per cent of the network, which is what keeps a model of this size practical to serve. It carries a 1-million-token context window and accepts text, images and video as input. Moonshot attributes the step up to two architectural changes it calls Kimi Delta Attention and Attention Residuals, paired with a framework it labels Stable LatentMoE. The company reports these yield roughly 2.5 times the scaling efficiency of its previous flagship, K2. Those efficiency figures are Moonshot's own and have not been independently reproduced.
On price, K3 sits at frontier rates rather than the discount tier some Chinese labs have used to compete. Moonshot lists $3.00 per million input tokens, $15.00 per million output tokens and $0.30 per million for cached input, billed flat across the full context window. That places it above open-weight rivals such as DeepSeek V4 and GLM-5.2, but generally below the published pricing of today's leading US closed-model APIs.
Where it lands on the benchmarks
On the independent Artificial Analysis Intelligence Index, a composite of nine benchmarks run the same way across models, K3 scores 57. At the time of writing, that places it third or fourth overall depending on the benchmark snapshot, behind only Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, as independent leaderboards continue to update when newly released models are evaluated. It ranks first on the Frontend Code Arena, a blind human-preference test for frontend code. Moonshot's own launch materials show it trading places with Claude Opus 4.8 and GPT-5.5 across coding and agentic tests, edging Opus 4.8 on Terminal-Bench 2.1 (88.3 to 84.6) while sitting just behind GPT-5.6 Sol at 88.8.
Those comparisons carry an important caveat. Each model was run inside its own tooling, K3 through Moonshot's KimiCode, the Claude models through Claude Code or Terminus 2, the GPT models through Codex, so the figures are not a like-for-like test and independent verification is still catching up. Moonshot's own model card also acknowledges that K3 lags Claude Fable 5 and GPT-5.6 Sol on user experience, an unusually direct admission in a launch document.
The release matters less for any single score than for where an open model now sits. Independent trackers place K3 within a few points of the best closed systems, narrowing a gap that analysts had put at six to nine months.
What it means for the region
For Southeast Asian teams the significance is practical rather than symbolic. Once the weights ship on 27 July, organisations with sufficient multi-node GPU infrastructure — or access to managed GPU cloud platforms — will have the option to self-host a model approaching frontier capability, and to fine-tune it on their own data rather than metering every call through a US provider's API. That lowers the entry cost of high-end capability for budget-constrained teams and aligns with the direction regional governments have already signalled through sovereign-model efforts such as Singapore's SEA-LION. Because the infrastructure bar is high, though, the near-term gain for most firms is cheaper hosted access and the option to customise, not literal on-premise control.
Key Takeaways
Kimi K3, released 16 July 2026, is a 2.8-trillion-parameter sparse-MoE model (16 of 896 experts active) with a 1-million-token context window and text, image and video input.
Full open weights are due 27 July; until then K3 runs only through Moonshot's hosted API and apps under the id
kimi-k3.Independent trackers place it third or fourth on the Artificial Analysis Intelligence Index and first on the Frontend Code Arena, behind only Claude Fable 5 and GPT-5.6 Sol.
Pricing is at frontier rates: $3.00 input, $15.00 output and $0.30 cached per million tokens, above open rivals but below leading US closed models.
Company-reported benchmarks were run in separate agent harnesses; independent verification and the licence terms are still pending.
Correction, 18 October 2026: An earlier version said Alibaba had announced a forthcoming 2.4-trillion-parameter open-weight Qwen model; we could not confirm that statement and have removed it. A disclosure has also been added.
View author profile → · Editorial policy
