Helpful links
The communities, tools, and docs behind everything on this site.
Local Inference Lab — benchmarks, builds & recipes
Independent open-source org for AI inference on hardware you control — and the source of most of what powers this site. Their llm-inference-bench is the best benchmarking tool for local LLM serving, and it produces the standardized run files used to import benchmark rows onto our leaderboard. Their container builds on Docker Hub — updated regularly — are the engines behind nearly every recipe here, including the Blackwell RTX PRO 6000 ones, and their Hugging Face org hosts checkpoints such as GLM-5.3-Flash-NVFP4. The rtx6kpro repo is where their Blackwell serving work lives. Results are posted in their Discord; releases ship with pinned tool versions so imported numbers stay attributable.
Website · llm-inference-bench (GitHub) · Releases · GitHub org · rtx6kpro — Blackwell recipe work · Discord · Docker launcher · Docker Hub — container images · Benchmarks site · Hugging Face · X
CMP Unlocker community — unlocked mining cards
The community around unlocking the Nvidia CMP 170HX — a mining card (CMP = crypto mining processor) built on an A100-class GA100 die and sold cheap for mining with most of its capability hidden behind mining firmware. Cards come in 64 GB and 80 GB variants, both HBM2e. The cmpunlocker tool patches the open NVIDIA kernel driver to restore what the firmware caps: full SM count, 64 GB of VRAM on 8 GB-profile cards (40 GB on 10 GB-profile ones), full BAR1, PCIe Gen 2, plus VFIO passthrough and profiling. Requires Linux, root, nvidia-open 610+, kernel headers, and Secure Boot off; the unlock survives reboots via patched modules. Discussions cover VBIOS flashing, power-limit tuning, cooling, and real serving results — the CMP 170HX machines behind the CMP recipes on this site come from this scene.
Artificial Analysis — model intelligence
Independent benchmarks of hosted and open models. Their AI Intelligence Index is the number shown in the AA column on our leaderboard (with a link to each model’s page and the index version we recorded). Models they do not score — mostly fine-tunes and vision experiments — show n/a rather than a made-up value.
Inference engines
The servers these recipes launch. vLLM powers every recipe on the site — PagedAttention, continuous batching, speculative decoding, and serves both OpenAI-compatible (/v1) and Anthropic-compatible (/anthropic/v1/messages) APIs. SGLang is the other engine llm-inference-bench auto-detects (radix prefix caching); recipes for it will land as they are verified.