FreshContextA context lab · Fife
Drawing
02 · Benchmarks
Sheet
2 of 3
Rev
A · 2026-10-01

MiniCPM5-2B · file← MiniCPM5-2B

STQ1_0 ternary

A ternary copy. Nothing could run it on a Mac's GPU, so it was never scored. Quantised here, llama.cpp-stq.

Size
0.7 GB
Status
On disk
Scored runs
0
ResultsNever scored

This file never completed a scored run. Its record below says why.

The recordHow it was made · how it was served · what went wrong · what happened to it

Rebuilt from the lab's records by Claude, an AI model. The results above come straight from the automated runs. How these pages are made.

FileMiniCPM5-2B-STQ1_0.gguf
FormatSTQ1_0 ternary, 700,728,416 B (2.21 bpw reported)
Where frommade here 2026-09-17, llama.cpp-stq llama-quantize with imatrix
Statuson disk

How it was made

  • Tool: llama.cpp-stq, upstream llama.cpp at 0cea36222 (the base named in the AngelSlim README) plus the STQ1_0 patch; Metal on, CUDA off; llama-quantize lists STQ1_0: 1.31 bpw ternarization.
  • Input: the BF16 GGUF with the imatrix made from it, both in a scratch folder (MiniCPM5-2B BF16 GGUF (converted here)).
  • Command:
llama.cpp-stq/build/bin/llama-quantize --imatrix minicpm5-2b.imatrix \
minicpm5-2b-bf16.gguf models/minicpm5-2b-stq1/MiniCPM5-2B-STQ1_0.gguf STQ1_0 8

The first attempt failed with ios_base::clear because the output directory did not exist.

  • Date: 2026-09-17 10:01Z. Output: quant size 663.36 MiB (2.21 BPW); 700,728,416 B on disk (stat, 2026-10-23). sha256 not on record.

How it was served

labeldates (UTC)binary and versionctx / -np / KV / -fa / --jinja / reasoningharness profilewhy this label exists
— (none)2026-09-17 10:02Z–10:29Z (CLI tries only)llama.cpp-stq/build/bin/llama-cli-ngl 999 crashed; -ngl 0 -t 8 too slow—never benched; no Metal kernel

Never served; no run-log label. -ngl 999 crashed in llama_decode (no Metal kernel); a CPU run with -ngl 0 -t 8 did not complete 16 tokens in 12 minutes and was killed.

Decisions

  • Verdict 2026-09-17 10:29Z: "the quantiser works on your Mac, the inference doesn't". Led directly to PTQ1_0.
  • Not published.

What went wrong

None beyond the failed first quantise (missing directory).

What happened to it

Made 2026-09-17 → never run → on disk at models/minicpm5-2b-stq1/ (Tier C leftover in a lab note). No RESTORE entry needed while held. Recreate with the command above.