MiniCPM5-2B · file← MiniCPM5-2B
Q8_0
Still on disk. OpenBMB's own Q8_0, our quality reference for the experiments. OpenBMB's GGUF.
- Size
- 2.7 GB
- Status
- On disk
- Scored runs
- 23
- Smallest Mac
- 8 GB estimate, 8K context
Results23 scored · 95% intervals
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| contractminicpm5-2b | 70.8% | 56.8–81.8 | Passed 34/48 | llama.cppextraction | 2026-09-15 |
| enrichmentminicpm5-2b | 59.3% | 46.0–71.3 | Passed 32/54 | llama.cppextraction | 2026-09-15 |
| enrichmentminicpm5-2b-q8-np1 | 57.4% | 44.2–69.7 | Passed 31/54 | llama.cppextraction | 2026-09-17 |
| epistemicsminicpm5-2b | 77.1% | 63.5–86.7 | Passed 37/48 | llama.cppstudy-nothink | 2026-09-15 |
| epistemicsminicpm5-2b-q8-np1 | 79.2% | 65.7–88.3 | Passed 38/48 | llama.cppstudy-nothink | 2026-09-17 |
| groundingminicpm5-2b | 62.4% | 53.7–70.4 | Passed 78/125 | llama.cppextraction | 2026-09-15 |
| groundingminicpm5-2b-q8-np1 | 56.0% | 47.2–64.4 | Passed 70/125 | llama.cppextraction | 2026-09-17 |
| nullschemaminicpm5-2b | 31.9% | 22.3–43.4 | Passed 23/72 | llama.cppextraction | 2026-09-15 |
| screeningminicpm5-2b | 85.0% | 73.9–91.9 | Passed 51/60 | llama.cppextraction | 2026-09-15 |
| screeningminicpm5-2b-q8-np1 | 83.3% | 72.0–90.7 | Passed 50/60 | llama.cppextraction | 2026-09-17 |
| synthesisminicpm5-2b | 60.0% | 47.4–71.4 | Passed 36/60 | llama.cppextraction | 2026-09-15 |
| synthesisminicpm5-2b-q8-np1 | 78.3% | 66.4–86.9 | Passed 47/60 | llama.cppextraction | 2026-09-17 |
| taggerminicpm5-2b | 51.7% | 45.4–57.9 | Passed 124/240 | llama.cppextraction | 2026-09-15 |
| taggerminicpm5-2b-q8-nojinja99.2% coverage | 76.9% | 71.1–81.8 | Passed 183/238 | llama.cppextraction | 2026-09-17 |
| taggerminicpm5-2b-q8-np1 | 76.7% | 70.9–81.6 | Passed 184/240 | llama.cppextraction | 2026-09-17 |
| taggerminicpm5-2b-q8-np4-today99.2% coverage | 76.9% | 71.1–81.8 | Passed 183/238 | llama.cppextraction | 2026-09-17 |
| tagger_nullableminicpm5-2b | 22.9% | 18.1–28.6 | Passed 55/240 | llama.cppextraction | 2026-09-15 |
| tagger_nullableminicpm5-2b-q8-np1 | 52.9% | 46.6–59.1 | Passed 127/240 | llama.cppextraction | 2026-09-17 |
| toolcallminicpm5-2b | 73.3% | 61.0–82.9 | Passed 44/60 | llama.cppextraction | 2026-09-15 |
| toolsminicpm5-2b | 93.3% | 70.2–98.8 | Passed 14/15 | llama.cppagentic-fast | 2026-09-15 |
| toolscaleminicpm5-2b | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppagentic-fast | 2026-09-15 |
| worldviewminicpm5-2b | 37.5% | 24.2–53.0 | Passed 15/40 | llama.cppextraction | 2026-09-15 |
| worldviewminicpm5-2b-q8-np1 | 57.5% | 42.2–71.5 | Passed 23/40 | llama.cppextraction | 2026-09-17 |
Runs marked unenforced were measured on a suite that requires constrained decoding, on a server shown not to apply the schema: mlx_lm accepts it and ignores it, and a few other servers and models break it. Those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.
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.
| File | MiniCPM5-2B-Q8_0.gguf |
| Format | GGUF Q8_0, 2,679,710,688 B |
| Where from | downloaded 2026-09-15 from openbmb/MiniCPM5-2B-GGUF |
| Status | on disk; benched; never seated |
| Hugging Face | https://huggingface.co/openbmb/MiniCPM5-2B-GGUF |
Where it came from
- Repo / file:
openbmb/MiniCPM5-2B-GGUF/MiniCPM5-2B-Q8_0.gguf(repo listing on 2026-09-15: F16, Q4_K_M, Q8_0). - Command:
hf_hub_download('openbmb/MiniCPM5-2B-GGUF','MiniCPM5-2B-Q8_0.gguf', local_dir='models/minicpm5-2b')viauv run --with huggingface_hub --with hf_transfer. - Date: 2026-09-15 18:48–18:50Z; file mtime 15 Sep 19:50 local (stat, 2026-10-23).
- Size: 2,679,710,688 B (stat, 2026-10-23). sha256 not on record. HF revision not on record.
- Header read at download: arch
llama, 42 layers, 2 KV heads, head_dim 128, ctx 131,072, 42.0 KB/token f16 KV. - Who asked: Alan, "Minicpm please".
How it was served
| label | dates (UTC) | binary | ctx / -np / KV / -fa / --jinja / sampling | harness profile | why this label exists |
|---|---|---|---|---|---|
minicpm5-2b | 2026-09-15 18:51Z–19:05Z | .local/bin/llama-server, b10893 (c32d1dabe) | -c 32768 -np 4 (8192/slot), --cache-ram 1024, -ctk q4_0 -ctv q4_0 -fa on -ngl 999, --jinja, --min-p 0 --repeat-penalty 1.15 (server-side; /props showed repeat_penalty 1.15 in force) | extraction; agentic-fast (toolscale, tools); study-nothink (epistemics) | the first 13-suite board, launched by a one-off script |
minicpm5-2b-q8-np1 | 2026-09-17 16:41Z–16:52Z | llama.cpp/build/bin/llama-server, b10893 | -c 8192 -np 1 -ngl 99 --jinja --no-warmup; f16 KV, default sampling (no min-p / repeat-penalty) | extraction; study-nothink | same-day paired baseline for EXL3, first suspected cause -np |
minicpm5-2b-q8-np4-today | 2026-09-17 16:53Z–16:57Z | same | -c 8192 -np 4 (2048/slot) --jinja | extraction | same server, same hour, only -np changed — killed the -np hypothesis |
minicpm5-2b-q8-nojinja | 2026-09-17 16:58Z–17:01Z | same | -c 8192 -np 4, no --jinja | extraction | tests the template path; also 76.9 |
The run-log llama_build (0.4.0-dev (build 10893, commit c32d1dabe)) matches the runtime for all four labels, so there is no stale stamp. No seat was ever held.
Unlabelled servers on this file: a thinking-off alias for EvalPlus; a stock control and a prism-fork control during the PTQ1_0 diagnosis; a coherence-gate control.
Decisions
- Downloaded as the 2026-09-15 board candidate (Alan). Served with the card's
--min-p 0 --repeat-penalty 1.15on purpose: "a stock launch runs MiniCPM5 with exactly the setting its card says breaks it". - Kept as the quality reference for the EXL3, STQ1_0 and PTQ1_0 experiments.
- Not seated: on 2026-09-20 rated "utility only; tagging is not its job", using the stale 2026-09-15 tagger numbers.
- Not published to HF (vendor file).
What went wrong
- Partial quoted as final. worldview 60.0% was reported mid-run; the final was 37.5%.
- Baseline not reproducible. tagger 51.7 (2026-09-15) vs 76.7 / 76.9 / 76.9 (2026-09-17); similar gaps on tagger_nullable, synthesis, worldview. Finding One baseline was 25 points off and nothing in the stamp could explain it; the other six reproduced exactly. What it cost: six control runs and a board spot-check. The fix was to stamp
model_path,model_ftype,model_alias,chat_template_sha. Not tested: the 2026-09-15 server's--repeat-penalty 1.15,--min-p 0, q4_0 KV and 8192-per-slot context, none of which the re-measures used or the manifests record (cache_type_k: nullthroughout) (reconstructed). - Stale numbers reused. The 2026-09-20 seat review quoted 51.7 / 22.9 after the finding said to re-measure first.
- Missing-flag check:
minicpm5-2b-q8-nojinjadeliberately ran without--jinja. Every other label had it.
What happened to it
Downloaded 2026-09-15 → benched 2026-09-15 and 2026-09-17 → never seated → on disk at models/minicpm5-2b/MiniCPM5-2B-Q8_0.gguf, listed as a Tier C leftover in a lab note. Re-downloadable from openbmb/MiniCPM5-2B-GGUF.