Moonshot's 2.8T Kimi K3 Drops, Open Weights July 27
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Moonshot AI released Kimi K3, a 2.8-trillion-parameter model with 1 million context length and native multimodal capabilities, claiming it rivals Anthropic's Opus 4.8 and GPT 5.5
- Kimi K3 is built on Kimi Delta Attention and Attention Residuals architecture with Mixture-of-Experts sparsity activating 16 of 2.8T parameters, achieving up to 6.3x faster decoding in million-token contexts and ~25% higher training efficiency
- Moonshot plans to release K3 model weights by July 27, pricing the model at $3/M input and $15/M output — the same price as Sonnet 5 while matching GPT-5.6 and Fable 5 on benchmarks
- Kimi K3 jumped to #1 on the Frontend Code Arena with 1679 points, a 17-place rise from Kimi-k2.6, claiming #1 in 6 of 7 frontend domains including Brand & Marketing and Data & Analytics
- Emad Mostaque estimates K3's total training compute at ~1e25 FLOPs at a cost of $15-$25M, per his public analysis of the model's specs
- Moonshot is reportedly fundraising at a $31.5B valuation, according to the Financial Times, in a round framed as challenging Anthropic's commercial lead
Why it matters: An open-weight model matching frontier benchmarks at Sonnet 5's price tier directly pressures Anthropic and OpenAI on both cost and openness — two dimensions neither currently competes on. The 17-place arena jump and sub-$25M training-cost estimate suggest the price disruption is backed by capability, not marketing.
