Topic

Moonshot AI (China) released Kimi K3 on July 16, 2026: a 2.8-trillion-parameter open-weight MoE model, the largest open model ever released. It scores above Claude Opus 4.8 on the GDPval-AA v2 real-world-tasks benchmark (1,668 Elo vs Opus 4.8's 1,600), while costing roughly half as much per task ($0.94 vs Opus 4.8). Full weights land July 27 under a Modified MIT license. This lands the same week as Gemini 3.5 Pro's rumored launch and Xi Jinping's appearance at Shanghai's World AI Conference, where 29 countries signed on to a new global AI cooperation body (WAICO).

Key facts used

• 2.8T parameters, Mixture-of-Experts, 1,048,576 token (1M) native context window with vision
• New architecture: Kimi Delta Attention (KDA) + Attention Residuals — KDA gives up to 6.3x faster decoding at million-token context
• GDPval-AA v2 (44 occupations, 9 industries): K3 scores 1,668-1,687 depending on source, ranking 3rd overall behind Claude Fable 5 Max (1,815) and GPT-5.6 Sol Max (1,748), ahead of Claude Opus 4.8 (1,600)
• Artificial Analysis independent eval: 57.11 Intelligence Index, 76.24 Coding Index, 50.07 Agentic Index
• Cost per task: ~$0.94, about half of Opus 4.8, though still pricier than smaller open-weight peers
• Pricing: $0.30/M cached input, $3/M input (cache miss), $15/M output at full 1M context
• Open-weight release under Modified MIT license, following the K2 pattern. Full weights ship July 27
• Released same week China hosted the World AI Conference in Shanghai; Xi called for AI development as "a symphony of international cooperation, not a solo performance"

Pillar

News drop + spicy opinion (China open-weight vs US frontier lab thesis). Continues the market-share pillar from Run 33 (Chinese models = 45% of OpenRouter dev traffic) and the compute/cost-war thread from Run 31. This is the sharpest version of that story yet: an open-weight model beating a closed frontier lab model on a real-world benchmark, not just a cost argument.

Post (LinkedIn)

A Chinese lab just open sourced a model that beats Claude Opus 4.8.

Moonshot AI released Kimi K3 on July 16. 2.8 trillion parameters. The largest open model ever shipped.

On GDPval-AA v2, the benchmark that tests real work across 44 occupations, K3 scores 1,668. Opus 4.8 scores 1,600.

It costs about half as much per task.

Full weights are free on July 27. Modified MIT license. Anyone can download it, fine-tune it, run it themselves.

This is not "China makes cheap knockoffs." This is China shipping frontier-level intelligence and giving it away.