Moonshot AI released Kimi K3 on July 27, and the number that matters is 2.8 trillion: total parameters in a mixture-of-experts design with 896 specialist experts, of which 16 activate per token for 104 billion active parameters. The context window is one million tokens, the model is natively multimodal, and the weights are on Hugging Face under moonshotai/Kimi-K3 in 4-bit MXFP4, which makes it the largest openly released model to date. On Artificial Analysis's Intelligence Index it lands third, behind Anthropic's Claude Fable 5 and OpenAI's GPT-5.6.
The pricing tells its own story. ChinAI's analysis of the launch calls the strategy affordable luxury: match the performance of the overseas flagships while anchoring the price to the ceiling of China's domestic market. K3's blended rate works out to about $2.30 per million tokens, with cached input at a tenth of the cache-miss rate, and its output price is 3.5 times higher than its predecessor K2.6. The domestic competition sits far below: Alibaba's Qwen3.7 Max at $1.40, Zhipu's GLM-5.2 at $0.90, DeepSeek V4 Pro at $0.18. The bet is that Chinese frontier labs can stop racing to the bottom, and early demand backed it in the least comfortable way possible: two days after launch, Moonshot suspended new subscriptions for lack of computing capacity and imposed usage limits.
The rest of the week was infrastructure. On July 29 Moonshot open-sourced MoonEP under MIT, the expert-parallelism communication library it credits with a 2.5x scaling-efficiency improvement on K3. Its core claim is exact load balance, every rank receives the same token count no matter how skewed the routing gets, which removes the per-layer host synchronization that throttles large MoE training; FlashKDA and AgentEnv shipped alongside it. On July 30, AWS published its official deployment guide, and it quietly defines what open means at this scale: the reference configuration is a single ml.p6-b300.48xlarge instance carrying eight NVIDIA B300 Blackwell Ultra GPUs, served through vLLM with tensor parallelism across all eight.
For builders, the week resolves into three facts. The largest open-weights model in history exists and you can download it today. Running it yourself starts at a rack of the newest accelerators NVIDIA sells, so open and accessible are not the same word. And the maker priced it like a flagship, not like a loss leader, then ran out of capacity anyway, which says more about where demand for frontier-class open models sits than any benchmark table does.