Remodeling LLM intelligence/cost charts: adding GLM-5.3 and local hosting costs
crusaderky · reddit · 2026-08-19
Criticizing Artificial Analysis' (AA) intelligence/cost plot for being misleading (specifically the use of a logarithmic scale for cost), the author created a linear-scale version. The new chart includes GLM-5.3, DeepSeek V4 Flash, and a self-hosted Qwen3.8-27B, with the latter's cost calculated based on electricity prices and RTX 3090 performance. The analysis suggests that for most users, the cost difference between running Qwen locally and using DeepSeek via API is negligible, though local hosting on dedicated high-end hardware (e.g., 128GB RAM) changes the economic equation significantly.
More from Infra
- FT warns of 'Darwin moment' for AI borrowers as CoreWeave bonds sink to distress levels — firstadopter · 2026-10-16
- Phone as interface, home servers as the lab: a two-workflow local AI setup — EAccelerate_42 · 2026-10-16
- "A datacenter is a bank": CoreWeave drama revives calls for算力 supervision — arthurcolle · 2026-10-16
- DOT-MoE: converting dense models into sparse MoE via optimal transport — burkov · 2026-10-16
- Stanford CS336: A 7B model needs 14GB for weights but 84GB to train — stanfordnlp · 2026-10-16
- New NVFP4 trick: compressing block scales saves 45-48% per expert matrix, bit-exact — TheZachMueller · 2026-10-16