Skip to content
RigCheck

How the numbers are made

A speed is labeled measured only when it comes from RigBase (CC BY 4.0) or from a submission a reviewer approved. RigBase aggregates public llama.cpp, MLX, Ollama, and similar reports and links each row to the original source. Speculative decoding (MTP, DFlash, and the rest) is stored but left out of the headline median. When several plain runs exist, the cell shows the median, the count, and the min–max. A source that printed a range stays a range. Nothing is filled in by scaling another GPU.

Weight memory is arithmetic. For a quantization with bits per weight b and P billion stored parameters:

weights_GiB = P × b / 8

The bits used for the three common columns:

  • Q4 · 4.85 bpw — Q4_K_M average, about 4.85 bits per weight. Q4 and Q8 follow the usual llama.cpp K-quant averages (llama.cpp). FP16 is 16 bits. MXFP4's 4.25 bits is the figure in OpenAI's gpt-oss model card, not a GGUF default.
  • Q8 · 8.5 bpw — Q8_0, 8.50 bits per weight. Q4 and Q8 follow the usual llama.cpp K-quant averages (llama.cpp). FP16 is 16 bits. MXFP4's 4.25 bits is the figure in OpenAI's gpt-oss model card, not a GGUF default.
  • FP16 · 16 bpw — 16 bits per weight. Q4 and Q8 follow the usual llama.cpp K-quant averages (llama.cpp). FP16 is 16 bits. MXFP4's 4.25 bits is the figure in OpenAI's gpt-oss model card, not a GGUF default.
  • MXFP4 · 4.25 bpw — 4.25 bits per weight, the figure OpenAI publishes for gpt-oss MoE weights. Q4 and Q8 follow the usual llama.cpp K-quant averages (llama.cpp). FP16 is 16 bits. MXFP4's 4.25 bits is the figure in OpenAI's gpt-oss model card, not a GGUF default.

Mixture-of-experts models use total stored parameters. Active parameters describe how much work a token does, not how big the file is. When a vendor publishes a checkpoint size, that size replaces the formula for that quant. gpt-oss MXFP4 is the case in this catalog.

KV cache is added only when the model card or the published config gives a full-attention shape. At 8,192 tokens, FP16 cache:

kv_GiB = 2 × layers × kv_heads × head_dim × 8192 × 2 / 1024³

Hybrid models count only the layers that store a KV cache. If the shape is not in the catalog, the KV term is omitted and the cell says so. A flat 0.75 GiB is then added for runtime overhead. The total is still an estimate.

Fit uses 90% of discrete VRAM and 70% of unified memory. "Tight" means the estimate is under the raw memory size but inside that reserve. A measured run is shown even when the estimate says the model does not fit. Offloading to system RAM is not scored.

Worked example, Llama 3.3 70B at Q4: weights 42.44 GiB, KV 2.50 GiB, overhead 0.75 GiB, total 45.69 GiB. That total is over a 24 GB card's usable memory.

gpt-oss 20B at MXFP4 uses the published checkpoint of 12.8 GiB plus 0.75 GiB overhead, with no KV term because the catalog does not store that shape. Total 13.55 GiB.

The weekly importer reads https://getrigbase.com/data/ and matches rows onto this catalog. Models without a sourced parameter count stay in the CSV and out of the picker.