Moonshot AI Launches Kimi K3: 2.8T Parameter Model
Original: Kimi K3, and what we can still learn from the pelican benchmark
Why This Matters
Kimi K3 raises the scale ceiling for open-weight models and signals aggressive pricing competition among frontier AI labs.
Chinese AI lab Moonshot AI announced Kimi K3 on July 16, 2026, describing it as their most capable model with 2.8 trillion parameters. Priced at $3/million input and $15/million output tokens, it is the most expensive model released by a Chinese AI lab to date. An open-weight release is promised by July 27, 2026.
Moonshot AI launched Kimi K3, positioning it as the first 'open 3T-class model,' surpassing DeepSeek's 1.6T v4 Pro. The model is currently available via the Moonshot website and API, with open weights promised by July 27, 2026. According to Artificial Analysis, Kimi K3 scored an Elo of 1547 on their private long-horizon knowledge work evaluation, a gain of 732 points over Kimi K2.6, ranking behind only Claude Fable 5. Cost per task is $0.94, roughly half the price of Claude Opus 4.8 at $1.80, and comparable to GPT-5.6 Sol at $1.04. The model also uses 21% fewer output tokens than K2.6. Kimi K3 currently leads Arena.ai's Frontend Code arena, surpassing Claude Fable 5. Pricing at $3/$15 per million tokens marks a significant jump from Kimi K2.6's $0.95/$4. Blogger Simon Willison ran his informal 'pelican benchmark'—generating an SVG of a pelican riding a bicycle—costing 25 cents due to heavy reasoning token usage. He notes the benchmark is now 21 months old and no longer reliably correlates to overall model quality, particularly as it does not evaluate agentic tool calling, which he identifies as the most critical capability for modern models.