A New Scale for Open Models
Kimi K3 activates 16 of 896 experts per token and ships with a one-million-token context window, native vision capabilities, and what Moonshot describes as “always-on” reasoning. The model is built on two architectural innovations the company calls Kimi Delta Attention, a hybrid linear attention mechanism that enables up to 6.3 times faster decoding at long contexts, and Attention Residuals, which selectively retrieves representations across model depth for roughly 25 percent greater training efficiency. youtube linkedin constellationr Cloudflare labellerr
In its own benchmarks, Moonshot says K3 “substantially outperformed” Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 while performing “competitively” with Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol, the two most capable proprietary models currently available. Independent evaluators corroborated strong results: Vals AI ranked K3 second of 38 models on its composite index at 74.7 percent, behind only Claude Fable 5. Artificial Analysis scored it at 57.11 on its Intelligence Index. On Arena.ai’s Frontend Code leaderboard, K3 claimed the top position, surpassing even Fable 5. artificialanalysis benchlm vals simonwillison vals fortune
Pricing and Open-Weight Timeline
The model’s API is live at $3 per million input tokens and $15 per million output tokens — roughly half the price of GPT-5.6 Sol and a third of Fable 5 on output, according to Moonshot’s published pricing. Full model weights are scheduled for public release by July 27 under a modified MIT license, which would make K3 downloadable and self-hostable. facebook instagram gmicloud labellerr
Simon Willison, a prominent AI developer, noted that K3 is “the most expensive model released by a Chinese AI lab to date” while still undercutting Western frontier pricing, and that the open-weight release would dethrone DeepSeek’s 1.6 trillion-parameter V4 Pro as the largest available open model. simonwillison
Funding and Strategic Context
The launch arrives as Moonshot seeks up to $2 billion in fresh capital at a $30 billion valuation, according to Bloomberg reporting from June, which would be the company’s third financing round in six months. Earlier this year Moonshot closed a $2 billion round led by Meituan’s venture arm that valued it above $20 billion, with Alibaba and Tencent among existing backers. Fortune reported that the release further narrows the gap between Chinese and U.S. models “just as global businesses are increasingly questioning the cost of deploying models from Anthropic and OpenAI.” japantimes Yahoo Finance bloomberg pulse2 fortune