Paper AI Tigers: Chinese frontier LLMs are 3-6x cheaper yet only 19% of OpenRouter use
gleech · x · 2026-10-16
gleech's essay "Paper AI Tigers" argues top Chinese LLMs offer frontier benchmark scores, 3x cheaper input / 6x cheaper output tokens, MIT-style open weights and visible CoT — yet outside China they're barely used: 19% of OpenRouter traffic combined and trending down through 2025. Martin Casado says 16-24% of US startups he meets use Chinese models; Airbnb uses Qwen, Windsurf's planner is likely GLM, Cursor's possibly DeepSeek. The author infers Chinese labs are severely compute-constrained (aggressive 32→4bit inference quantization), so real capability likely trails US flags. In the thread he also cautions against trusting Meta's new model's benchmarks: predecessor Muse Spark 1.3 settled around #15 on OpenRouter (Kimi K3 level), and Wonder is expected around #12.
More from Models
- NVIDIA Nemotron 3.5 Super VL lands in Perplexity for Pro/Max and Agent API — perplexity_ai · 2026-10-16
- Meta's 816B open-weight Muse Spark Open hits Groq preview at 350+ tokens/sec — GroqLLC · 2026-10-16
- Meta's 816B Muse Spark Open runs locally on two Mac Studios via llama.cpp — mervenoyann · 2026-10-16
- User evals: Opus 5.5 still leads on price-quality-speed, Sol for daily use — dreamwieber · 2026-10-16
- Meta ships Muse Spark Open: 816B MoE with 1M context, runs on two M5 Ultras — awnihannun · 2026-10-16
- First hands-on with Gemini 4 Argon: standout writing beats Opus 5.5 and GPT 6 Astra — soumitrashukla9 · 2026-10-16