Written by AI agents, curated and verified by me.
Kimi K3: open weights in the 3T class, and nothing verified yet
- Kimi
- Coding Agents
- Verification
Moonshot AI has introduced Kimi K3, by its own account the first open model in the 3T class: 2.8 trillion parameters, a one-million-token context window, native vision. The full weights are due by 27 July 2026. The notable part is not a position in a ranking. It is that a model of this size will soon exist as a file, and that nearly every coding figure attached to it comes from the vendor.
What is Kimi K3?
Kimi K3 is a mixture-of-experts model with 2.8 trillion parameters, of which, according to Moonshot AI, 16 out of 896 experts are effectively active. Architecturally the team names Kimi Delta Attention and Attention Residuals plus a Stable LatentMoE framework, together about 2.5 times better scaling efficiency than Kimi K2. At launch the model runs with max thinking effort by default; low- and high-effort modes are to follow. It is available today in the Kimi apps, in Kimi Work, through the API as kimi-k3, and in the terminal: in Kimi Code you select it with /model. The API costs 0.30 dollars per million input tokens on a cache hit, 3.00 dollars on a cache miss, and 15.00 dollars per million output tokens.
What do the coding numbers actually say?
Less than the headline suggests. Moonshot AI itself places K3 behind Claude Fable 5 and GPT 5.6 Sol, claiming the lead only over the other tested models. The prominent margin over Claude Opus 4.8 comes from a case study on GPU kernel optimisation: four tasks, up to 24 hours per model in an identical sandbox. There, Moonshot reports K3 substantially ahead of Opus 4.8, GPT 5.6 Sol, and GPT 5.5, and competitive with Fable 5, though Fable 5 was evaluated by a third party and ran into fallback behaviour. On DeepSWE, K3 reaches 67.3 with the mini-SWE-agent harness per the official leaderboard. The footnotes carry the part that matters: K3 is often measured with Moonshot’s own KimiCode harness while the comparison models run under Claude Code or Codex, in places on different hardware. Numbers like these are indications, not evidence. What comes out on your tasks is decided by your own test run.
What do open weights mean at this size?
For most teams, less self-hosting than the word implies. Moonshot AI recommends deployment on supernode configurations with 64 or more accelerators. If you do not have that in house, you use the API or wait for hosters. The value sits elsewhere. A model whose weights are open can be inspected, reproduced, and evaluated independently by third parties. That is exactly what every proprietary benchmark table lacks. From 27 July, the coding claims in this announcement can be re-measured by people who gain nothing from them being true. That is the difference, not the score.
Which limits does Moonshot name itself?
Two, and both hit agent operations directly. First, K3 is sensitive to thinking history: if the harness does not pass back the full historical thinking content, or if you switch an ongoing session from another model to K3, generation quality can become highly unstable. Moonshot advises using a harness with verified compatibility and not switching mid-session. Second, the team describes “excessive proactiveness”: on minor issues or ambiguous intent, K3 tends to decide on your behalf rather than ask. The advice is to set boundaries explicitly in the system prompt or in AGENTS.md. Moonshot also concedes a noticeable gap in user experience against Fable 5 and GPT 5.6 Sol. The candour is welcome, but the consequence stands: a model that likes to decide on its own needs tighter guardrails, not looser ones.
Where Kimi K3 earns its place
On long, tedious stretches with a lot of context: large repositories, terminal work, research across thousands of pages. On BrowseComp, Moonshot reports 90.4 when the full context window is used without compaction. That is precisely the class of task where an agent earns its place. The decisions with consequences you still pull back out: before the merge, before the deploy, before the migration. With Kimi K2.7 Code the observation was already the same, open weights lower the barrier, not the responsibility. At 2.8 trillion parameters that holds unchanged, and as described in agentic engineering, reliability comes from the architecture around the model. Until 27 July, this announcement remains a claim with a delivery date.