Open-Source Benchmarks: RTX 5090 LLM Quants and 8GB VRAM Agentic Scores

max_paperclips · x · 2026-08-06

Developer witcheer has open-sourced a comprehensive suite of benchmark data and tools for local LLM deployment, including:

Hardware & Quantization: Detailed speed and quality comparisons for GGUF quants (Q8 down to Q3) on the new RTX 5090.
Low VRAM Limits: Tested 11 models on an 8GB RTX 4060 Ti, evaluating speed, quality, and agentic capabilities.
Local Agent Eval: Released a leaderboard of local models capable of driving agents, plus an 8GB VRAM edition agentic coding bench.
Tools: Open-sourced llm-bench-rig, the dual-engine (llama.cpp + vLLM) pipeline powering these tests, along with tested Hermes recipes.

More from coding & agent