Moonshot’s Kimi K3 Becomes Largest Open-Weight AI Model Ever

BEIJING: Moonshot AI releases Kimi K3, a 2.8 trillion-parameter model that instantly becomes the largest open-weight system ever shipped.

The July 16 launch puts K3 live on Kimi’s apps and API immediately. Full open weights follow by July 27. Moonshot calls it the first open 3T-class model, rounding up with visible pride.

The specs read like a closed frontier model’s. Mixture-of-experts architecture, a one-million-token context window, native multimodal understanding, and reasoning switched on by default. No separate thinking variant. Two versions ship at launch: K3 Max for chat and agents, K3 Swarm Max for parallel processing.

Self-reported benchmarks place K3 above everything except Fable 5 and GPT-5.6 Sol. It beats Claude Opus 4.8 and GPT-5.5 on Moonshot’s tables, and independent testing so far lands it fourth among all frontier models. It even took the Frontend Code Arena crown from Claude Opus 5.

Pricing is the sharper weapon. K3 costs $3 per million input tokens and $15 for output. Fable 5 charges $50 for the same output. Expensive by Chinese standards, where DeepSeek V4 runs at 87 cents, yet a fraction of the American equivalent.

Silicon Valley already runs on Kimi more than it admits. Cursor used it to help build Composer 2. DoorDash delegates routine work to K2.6. Thinking Machines tapped K2.5 for post-training data on its new Inkling model.

Bank of America analysts note that “pre-training scaling, paired with architectural innovation, can still deliver step-change gains” despite China’s compute constraints. Tom’s Hardware reports training ran on export-grade Nvidia silicon plus an unnamed alternative GPU vendor.

Moonshot raised $2 billion in May at a valuation above $20 billion. Washington is reportedly weighing how to slow American adoption of Chinese models. K3 just made that conversation harder.