Model components / Base / pretrained models

Kimi K3

Moonshot AI's flagship 2.8T-parameter MoE (activating 16 of 896 experts) with native vision and a 1M-token context window, announced 2026-07-16. Positioned as the world's first open 3T-class model; live on kimi.com / Kimi Work / Kimi Code / the Kimi API at launch, with full weights scheduled for 2026-07-27. Built on Kimi Delta Attention and Attention Residuals; Moonshot reports frontier-tier long-horizon coding and agentic knowledge-work results trailing only Claude Fable 5 and GPT-5.6 Sol among tested models.

Kimi K3, 2.8T total MoE, Stable LatentMoE (16/896 experts), 1M context, native multimodal. API live 16 Jul 2026; weights promised by 27 Jul 2026. Openness scored assuming continuity with the Kimi K2.x Modified-MIT open-weight license (same family pattern as kimi-k2-6); confirm against the HF card once weights land. No HF URL yet — add when the checkpoint is published. Consolidated on 2026-07-29 from kimi-k2-6; openness follows kimi-k3, the release that currently governs. See sources/slug_aliases.yaml.

Openness

3 medium confidence

3.0

weights: pending(scheduled 2026-07-27
data: closed
code: partial(inference/deploy expected
license: assumed-Modified-MIT(continuity with Kimi K2.x

Modified-MIT weights - standard MIT below a 100M-MAU and $20M-monthly-revenue threshold - with a closed corpus, so 3. Confidence is medium rather than high on purpose: K3 governs as the current release, and its license is recorded as `assumed-Modified-MIT` by continuity with K2.x rather than confirmed against a published file. K2.6's Modified-MIT is confirmed, so the tier's openness firms up as soon as K3's terms are read directly. Adoption 4 is carried from K2.6.

https://www.kimi.com/blog/kimi-k3 recorded 2026-07-19
primary announcement — 2.8T MoE, 1M context, native vision; API live; full weights by 2026-07-27
https://huggingface.co/moonshotai/Kimi-K2.6 recorded 2026-07-19
prior Kimi open-weight release under Modified MIT (assumed continuity for K3 license class)

Adoption

4 high confidence

4.0

Free on kimi.com + Kimi app + API + Kimi Code; partners Vercel, Baseten, Ollama, Factory.ai validated K2.6 in production (Vercel reports >50% improvement on Next.js benchmark). OpenRouter shows sustained post-launch usage spikes indicating genuine production adoption. No exact download/user count published; placed at 4 (1-10M-equivalent) on production-traction signal.

https://www.marktechpost.com/2026/04/20/moonshot-ai-releases-kimi-k2-6-with-long-horizon-coding-agent-swarm-scaling-to-300-sub-agents-and-4000-coordinated-steps/ recorded 2026-06-04
flagship phase-C verification source
https://blog.kilo.ai/p/kimi-k26-has-arrived-an-open-weight recorded 2026-06-04
flagship phase-C verification source

Capability

5 high confidence

5.0

SWE-bench Verified 80.2%, SWE-bench Pro 58.6% (up from 50.7% in K2.5), Terminal-Bench 2.0 66.7% (up from 50.8%). Reported tying GPT-5.5 on coding and beating some top US models; specialized for long-horizon agentic coding (300-agent swarms, 4,000 coordinated steps). Frontier-tier for agentic/coding work.

https://benchlm.ai/models/kimi-2-6 recorded 2026-06-04
flagship phase-C verification source
https://officechai.com/ai/kimi-k2-6-benchmarks/ recorded 2026-06-04
flagship phase-C verification source
https://www.kimi.com/blog/kimi-k2-6 recorded 2026-06-04
flagship phase-C verification source

Unchanged since 2026-07-30 (last edited, not re-checked)