Moonshot AI Unveils Kimi K3: A 2.8 Trillion Parameter Model with High Costs
Chinese AI lab Moonshot AI has launched Kimi K3, its most advanced model to date, boasting 2.8 trillion parameters. While currently accessible via their website and API, an open-weight release is anticipated by July 27, 2026. Moonshot AI claims Kimi K3 is the first "open 3T-class model," surpassing DeepSeek's 1.6T v4 Pro. Self-reported benchmarks indicate Kimi K3 generally outperforms Claude Opus 4.8 and GPT-5.5 high, though it falls short of Claude Fable 5 and GPT-5.6 Sol. The model shows significant improvement over its predecessor, Kimi K2.6, with a substantial increase in its Elo score on long-horizon knowledge tasks. However, Kimi K3's operational costs are notably high, priced at $3 per million input tokens and $15 per million output tokens. This pricing aligns with Anthropic's Claude Sonnet series and marks the most expensive offering from a Chinese AI lab to date, a considerable jump from Kimi K2.6's $0.95/$4 rates. Despite the cost, Kimi K3 demonstrates efficiency in token usage, consuming 21% fewer output tokens than Kimi K2.6, and leads on Arena.ai's Frontend Code arena. The model also exhibits strong vision capabilities, accurately interpreting an SVG image and generating detailed alt text.
The release of Kimi K3 by Moonshot AI highlights a competitive landscape in large language model development, characterized by escalating parameter counts and sophisticated performance claims. The significant increase in parameter count and associated pricing suggests a strategic market positioning, potentially targeting enterprise-level applications where high capability justifies premium costs. This move by a Chinese AI lab also underscores the growing global influence and technological parity in advanced AI research. The substantial cost per token, while high, is presented alongside efficiency gains and competitive benchmark results, prompting an examination of the value proposition for users. Future developments will likely focus on balancing computational expense with demonstrable utility, particularly in areas like agentic tool calling, which remain critical for practical AI deployment beyond benchmark performance.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.