MiniCPM5-2B · file← MiniCPM5-2B
BF16 safetensors
OpenBMB's full-size original. We made the experimental copies from it, then deleted it. OpenBMB.
- Size
- 4.7 GB
- Status
- Deleted 11 Oct 2026
- 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.
| File | minicpm5-2b-hf/ (safetensors + tokenizer/config) |
| Format | BF16 safetensors, 4.7 GB (du) |
| Where from | downloaded 2026-09-16 and again 2026-09-17 from openbmb/MiniCPM5-2B |
| Status | deleted 2026-10-11 (earlier deleted 2026-09-16 20:10Z and re-downloaded 2026-09-17 09:55Z) |
| Hugging Face | https://huggingface.co/openbmb/MiniCPM5-2B |
Where it came from
- Repo:
openbmb/MiniCPM5-2B. Command both times:
snapshot_download('openbmb/MiniCPM5-2B', local_dir='models/minicpm5-2b-hf',
allow_patterns=['*.safetensors','*.json','*.txt','*.jinja','*.model'], max_workers=4)- Dates: 2026-09-16 13:37–13:41Z (8 files, 4.7 GB du) and 2026-09-17 09:55–09:59Z.
- Size: 4.7 GB (du); 4.8 GB per a lab note. Byte count, sha256 and revision: not on record.
- Who asked: the 2026-09-16 pull followed Alan's "try it, am at work, give you something to do :)" on the EXL3 question. The 2026-09-17 pull was for the ternary test that followed Alan's "Can you did ruthlessly online to see if there’s any way forward with inference" — the STQ1_0 attempt preceded that message by two hours, so its own trigger is not on record.
- The HF cache also held
models--openbmb--MiniCPM5-2Bat 3.6 GB on 2026-10-11.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / reasoning | harness profile | why this label exists |
|---|---|---|---|---|---|
| — (none) | — | — | — | — | never served; conversion input only |
Never served; no run-log label. It was conversion input only.
Decisions
- Downloaded because no GGUF path existed for PonyExl3.
- Deleted 2026-09-16 20:10Z to make room for DeepSeek-V4-Flash (Alan: "Go for it then").
- Re-downloaded 2026-09-17 09:55Z as the BF16 source for STQ1_0/PTQ1_0.
- Deleted 2026-10-11 in the third sweep pass (a lab note).
What went wrong
- The 2026-10-11 sweep note says "only the GGUF (minicpm5-2b/) was ever benched". True of this folder, but the EXL3 build made from it was benched (11 run-log entries); see the EXL3 variant doc.
What happened to it
Downloaded 2026-09-16 → EXL3 source → deleted 2026-09-16 → re-downloaded 2026-09-17 → BF16 GGUF source → deleted 2026-10-11 (the restore script lists it but has no download command). Re-downloadable from openbmb/MiniCPM5-2B with the command above.