Skip to content- Watching for: The 2026-07-27 weight release and license (K2 was modified-MIT); the technical report; independent benchmark reproduction; whether the $3.00/$15.00 per 1M pricing holds as third parties serve the open weights; whether third-party inference providers absorb the demand behind the subscription pause; the full UK AISI/CAISI report and methodology; whether the safeguards finding draws a Moonshot response.
Notes
- 2026-07-25: Moonshot published the Kimi K3 announcement 2026-07-16, adding detail missing at go-live: 2.8T-param MoE (Stable LatentMoE, 16 of 896 experts active/token), Kimi Delta Attention, 1M context, native multimodal. API model id
kimi-k3, $0.30 cache-hit / $3.00 cache-miss per 1M input, $15.00 per 1M output. Full weights promised by 2026-07-27, technical report to follow. Vendor figures rank overall intelligence behind Fable 5 and GPT-5.6 Sol; HN notes the pricing is high for a Chinese open-weight model. Simon Willison calls it the first "open 3T-class model". Covered 2026-07-17 Top stories. Supersedes the 2026-07-16 "Kimi K3 live without a model card" note. 2026-07-18: K3 reached #1 in the Frontend Code Arena. 2026-07-19: Moonshot (@kimi_moonshot) announced it is suspending new subscriptions on K3 demand, existing subscribers keep access; HN reports K3 strong for code/PR review but slow under load. Covered 2026-07-20 Top stories (confirmed). 2026-07-23: UK AI Security Institute and US CAISI published a joint preliminary cyber-capability assessment via NIST (nist.gov, HN 49044492): ExploitBench 32% over 41 post-2023 V8 CVEs (GLM 5.2 24%, below recent frontier models), arbitrary code execution on 0 of 41 tasks against a leading-model average of 20 of 41; The Last Ones 32-step simulated network intrusion reached step 17 on average against 28.5 for US frontier models tested with safeguards disabled, full completion 1 of 10 attempts; K3 safeguards did not prevent exploit-development attempts. Stated preliminary over a small benchmark set; the network scenario has no active defenders and an intentional attack path. A government measurement bounding the capability behind the Redis zero-day claims. Covered 2026-07-25 Top stories (Security, confirmed).
Stories