GLM-4.7-Flash · file← GLM-4.7-Flash
bf16 safetensors
Z.ai's full-size original. Never run; we made the four MLX builds from it. Zai-org.
- Size
- 62.5 GB
- Status
- Deleted 9 Sep 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 | glm-4.7-flash-bf16/ (48 safetensors shards) |
| Format | bf16, 58.2 GiB (HF sum) |
| Where from | downloaded 2026-09-07/08 from zai-org/GLM-4.7-Flash |
| Status | deleted 2026-09-09 14:08Z |
| Hugging Face | https://huggingface.co/zai-org/GLM-4.7-Flash |
Where it came from
- Repo:
zai-org/GLM-4.7-Flash, all files, unpinned revision (revision not on record). - Command (resumed by a one-off script):
export HF_HUB_ENABLE_HF_TRANSFER=1 HF_XET_HIGH_PERFORMANCE=1
$HF download zai-org/GLM-4.7-Flash --local-dir glm-4.7-flash-bf16HF was bfcl-venv/bin/hf in the resume script.
- When: started 2026-09-07 23:18 BST (22:18Z). Resumed 00:03 BST. Complete before the 00:38 BST conversion.
- Size: 48 safetensors shards, 58.2 GiB summed from the HF listing;
du58–60 G on disk. Exact bytes and sha256 not on record. - Who asked: Alan, "Can we quantize it?".
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
| — | — | — | — | — | never served. It was the input to mlx_lm convert, mlx_lm.dynamic_quant (failed) and the deferred AWQ/GPTQ |
No seats held.
Decisions
- Downloaded 2026-09-07 to make the first MLX builds of the model.
- Kept 2026-09-08 07:50 BST. Alan: "We should try mixed precision etc so keep the bf16 GLM". Kept again at 14:29Z for AWQ/GPTQ, after Alan's "we would be fools to not take it all the way with awq and gptq".
- Deleted 2026-09-09 14:08Z. Inventory: "conversion source, job done". Alan: "ok remove the granite extras and the superseded models".
- Not published (it is upstream's file).
What went wrong
- A 58 GB load on a 64 GB box.
convertworked because it streams shard by shard; AWQ/GPTQ need the weights resident, which is why they were parked. mlx_lm.dynamic_quantagainst this source failed on the MoE gather gradient, not memory.
What happened to it
Downloaded 2026-09-07/08 → converted into four MLX builds 2026-09-08 → deleted 2026-09-09 14:08Z. No RESTORE entry. Re-downloadable from zai-org/GLM-4.7-Flash, but a later revision may differ from the one converted here, which was not pinned.