Vitamin K3
K3
Vitamin K3 is a large language model developed by Moonshot AI, better known in the market as **Kimi K3** . It is a 2.8 trillion-parameter model built on KDA (Kimi Delta Attention) hybrid linear attention and Attention Residuals, with native vision understanding and a 1 million token context window . Moonshot AI released it in July 2026 as the world's first open-source model at the ~3 trillion parameter scale, targeting long-horizon programming, knowledge work, and reasoning .
In benchmark terms, K3 ranks just behind the top closed-source models Claude Fable 5 and GPT-5.6 Sol overall, but leads in specific domains: it topped Arena AI's Code Arena for front-end programming and scored first in front-end development, legal work, spreadsheets, creative writing, and web development in broader evaluations . At launch, it was priced at $3 per million input tokens and $15 per million output tokens—roughly 40% of Claude Opus and 20% of Claude Fable costs . Moonshot AI also open-sourced the full weights, technical report, and training infrastructure on July 27, 2026 . The model was still in training as of early 2026, with Moonshot AI's $500 million Series C round that January earmarked for accelerating K3 development .
AI-generated — may contain errors, please verify.
Coverage
Open-source models are the united front
War of Attrition
Only second-rate models do anonymous benchmark bragging
What the hell is the whole bull market coming for if the big models can't even guess your horse? --- *Note: This appears to be Chinese internet slang/meme language. "牛来" (niú lái) = "bull market coming" (from crypto/stock trading slang). "猜你马" (cāi nǐ mǎ) is a vulgar homophone of "猜你妈" (guess your mom), used here to express frustration that AI models can't predict things accurately — roughly equivalent to "what's the point" or "this is useless." The tone is crude and exasperated, typical of Chinese social media shitposting.*
Large Language Models Are a Brutal Manufacturing Business
Get Yu Hao to train the model, stat.
Come to the Internet Cafe and Put Every LLM to the Real-World Test
Everyone's an AI Model Whisperer Now
Big Models Can Work, But Dongmodi Knows How to Fly
Which Browser Is Best for Running AI Models?
Qwen 3.8 Upgrades to Pure "Digital Laborer" Model
Don't be Wang Feng next time.
Meituan's LLM Is as Free as Its Delivery Service
Bullish on Meituan building world models
5Y News|Moonshot AI Closes $500 Million Series C Round, Aims to Become a World-Leading AGI Company
Moonshot AI will become an LLM that is "distinctive" and "undefinable."
Moonshot AI Completes $500 Million New Funding Round, Zhilin Yang: Kimi Will Become a Distinctive LLM
"Become the world's leading AGI company."








