Kimi K3: Open Frontier Intelligence

1822 points · 1069 comments on HN · read original →

Points and comments are a snapshot, not live.

Kimi releases K3, a 2.8T-parameter open-source model rivaling proprietary frontiers.

Kimi K3 is a 2.8 trillion parameter open model with 1M-token context and native vision, using Kimi Delta Attention and Attention Residuals. It trails Claude Fable 5 and GPT 5.6 Sol but outperforms other tested models on coding, knowledge work, and reasoning benchmarks. It activates 16 of 896 MoE experts and uses MXFP4 weights. Pricing is $3/MTok input, $15/MTok output. Full weights will release by July 27, 2026.

The model demonstrated long-horizon coding by optimizing GPU kernels, building a Triton-like compiler, designing a chip, and reproducing astrophysics research. It also handles game dev, video editing, and interactive visualizations for financial and scientific research.

What commenters are saying

Commenters focused on pricing and real-world cost efficiency. Several noted Kimi K3's $3/$15 per MTok pricing matches Anthropic's Sonnet series and is close to GPT 5.6 Terra, calling it high for a Chinese open-weight model but justified if truly near frontier capability. A key concern was reasoning token efficiency: if K3 spends far more reasoning tokens than GPT or Claude on a task, its higher token burn negates the per-token price advantage. Some users shared experiences that K2.7 and GLM models were token-inefficient in practice, making subscriptions or cheaper alternatives like DeepSeek V4 Flash better value for daily coding.

Two camps emerged: those arguing the high price is justified by capabilities rivaling Fable/Sol, and those warning that poor reasoning efficiency and tokenizer differences make comparisons deceptive. Several commenters noted potential for inference providers to undercut official pricing once weights are open.