Public wiki / concepts / ai / models

Pages

• Comparing LLMs by coding benchmark — What common coding benchmarks (SWE-bench, LiveCodeBench-style aggregates, arena ratings) actually measure and how to read a price/performance leaderboard.
• KV-cache cost — How to empirically determine the safe max context window for a local LLM based on measured VRAM cost per token.
• Current LLM coding leaderboard — A refreshed top-100 coding leaderboard combining BenchLM coding scores with OpenRouter token pricing.
• llama.cpp vs Ollama on an older Pascal datacenter GPU — How llama.cpp compares to Ollama for throughput and KV-cache efficiency on a previous-generation (Pascal) datacenter GPU.