Initial release: 117 builds, 298 measured results (2026-10-24)

README.md

@@ -1,3 +1,59 @@
 ---
 license: cc-by-4.0
+pretty_name: RigBase - measured local LLM speeds on home AI hardware
+language:
+- en
+tags:
+- llm
+- benchmark
+- gpu
+- local-llm
+- llama.cpp
+- hardware
+- tokens-per-second
+size_categories:
+- n<1K
+configs:
+- config_name: default
+  data_files: builds.csv
 ---
+
+# RigBase: measured local-LLM speeds on real home hardware
+
+117 real home AI builds and 298 measured results across 48 models (Qwen3.8, Qwen3.6, Gemma 4, gpt-oss, Nemotron, DeepSeek and more), on NVIDIA, AMD, Intel, Apple Silicon, Strix Halo and DGX Spark. Snapshot: 2026-10-24.
+
+Every row links to its original public source (GitHub discussions, Reddit, llama-bench databases, blogs, YouTube). RigBase collects and normalises these numbers; it does not re-measure them.
+
+- Website: https://getrigbase.com
+- Always-current download: https://getrigbase.com/data/ (builds.csv, builds.json)
+- Submit a build: https://getrigbase.com/submit/
+
+## Files
+
+- `builds.csv`: one row per measured result (298 rows).
+- `builds.json`: one object per build with its parts and a list of results.
+
+## Columns (builds.csv)
+
+| column | meaning |
+|---|---|
+| build_slug, build_title | build id and name; page at page_url |
+| gpus, vram_gb | GPU(s) or chip, total VRAM / unified memory in GB |
+| cpu, ram, power | host parts and reported power draw (free text) |
+| model, quant, engine | model, quantization, inference engine as reported |
+| tps | generation (decode) speed, tokens/sec |
+| spec_decode | `none`, or the speculative-decoding method (MTP, draft model, etc.) |
+| prefill_tps, context | prompt-processing speed and context length, where reported |
+| how_measured, status | measurement notes and status |
+| source_checked_by_rigbase | whether RigBase re-checked the number against the source |
+| source | URL of the original public measurement |
+| page_url | RigBase page for the build |
+
+## Caveats
+
+- Numbers come from different people, engines, quants and settings. Compare rows from the same source when you compute ratios.
+- Results with speculative decoding are marked in `spec_decode`; filter them out for plain decode speed.
+
+## License and citation
+
+CC BY 4.0. Please credit "RigBase (getrigbase.com)" and link the original sources where you use individual numbers.