Kimi K3
The world's first open 3T-class model — 2.8T parameters of always-on reasoning for long-horizon coding agents.
About
Kimi K3 is Moonshot AI's flagship: a 2.8T-total-parameter MoE (896 experts, 16 active per token) with a novel hybrid linear/full attention design, native vision, and thinking that is always on (reasoning_effort low/high/max). Announced July 16, 2026 with open weights released July 26–27 under Moonshot's own Kimi K3 License.
It is explicitly built for long-horizon agentic coding and knowledge work, with day-zero vLLM/SGLang support.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Independent: Artificial Analysis Intelligence Index 57, ranked 4th of 189 models (via DataNorth).
Moonshot-reported: first place on two internal coding tests; #1 on the Frontend Code evaluation (1679).
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 9/10
Strong independent benchmark corroboration (AA Index 57, 4th of 189) and genuinely open weights with day-zero vLLM/SGLang support. Caveats: the Kimi K3 License is bespoke rather than a standard permissive license, Moonshot's docs and community support are thinner than Western labs', pricing is mid-tier rather than cheap, and there is no verified free tier.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.9/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 7.2/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · HF top match: RadixArk/Kimi-K3-DSpark (community upload counts as proxy where official repo is gated)
- · HF downloads 5,724,892 (peer-normalized)
- · released 19 days ago
- · manual review: thin English docs; no independent evals yet