Kimi K3 Open Weights July 27, 2026: What to Know
Kimi K3 Open Weights July 27, 2026: What to Know
Moonshot AI’s Kimi K3 — a 2.8-trillion-parameter MoE model — gets its full open weights on July 27, 2026, becoming the strongest open model yet released. Here’s what it is, how it benchmarks, and why it matters.
Last verified: July 25, 2026
What Kimi K3 Is
Moonshot AI announced Kimi K3 on July 16, 2026 with immediate API access and a promise of open weights on July 27, 2026. Key specs:
- 2.8 trillion parameters, Mixture-of-Experts (sparse activation)
- 1 million-token context window
- Native vision (multimodal)
- Flat $3/$15 per MTok API pricing (no peak/valley split)
How It Benchmarks
Kimi K3 is the headline because of where it lands on the leaderboard, not just that it’s open:
- ~57 on the Artificial Analysis Intelligence Index — ranked #3 overall, behind only Claude Fable 5 and GPT-5.6 Sol, and comparable to Opus 4.8 and GPT-5.5.
- Compared to other open trillion-scale MoE models, it beats DeepSeek V4 Pro (44) and GLM-5.2 (51) on that index.
- Independent comparisons rate it “more intelligent” than DeepSeek V4 Pro with a stronger capability profile and native vision.
This is the first time an open-weight model has cracked the very top tier of the intelligence index — previously the exclusive territory of US closed frontier labs.
Pricing and Cost per Task
At a flat $3/$15 per MTok, a real 30K-input/5K-output task costs about $0.165 — roughly half the cost of GPT-5.6 Sol ($0.30) or Claude Opus 4.8 ($0.28) for comparable frontier-tier intelligence.
Can You Self-Host It?
Once weights land on July 27, self-hosting is technically possible — but a 2.8T sparse MoE still needs distributed infrastructure. Viral “run it for $600M” and “Microsoft-hosted inference” scenarios circulated in July, but running K3 locally is realistic only for teams with serious multi-GPU clusters. Most builders will use the API or a hosted inference provider. For everyday local use, smaller open models (DeepSeek V4 Flash, GLM, Qwen dense variants) remain the practical choice.
Why It Matters
Kimi K3 escalates the open-weights race: an open model now sits inside the frontier tier, at half the price of US flagships, with a permissive-enough release to build on. For teams that want frontier quality without vendor lock-in — or that need on-prem for compliance — K3 changes the calculus.
Bottom Line
- Open weights: July 27, 2026
- Capability: #3 overall (~57 index), top open model ever
- Price: flat $3/$15, ~$0.165/task — half of US flagships
- Catch: 2.8T MoE needs distributed infra to self-host