GLM-4.7-Flash · file← GLM-4.7-Flash
MLX mixed_3_6
A mixed-precision MLX build, the best for its size. Meant for upload, never uploaded. Made here from bf16.
- Size
- 14.9 GB
- Status
- Deleted 16 Sep 2026
- Scored runs
- 1
- Smallest Mac
- 32 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| screeningglm-4.7-flash-mlx-mixed_3_6 | 93.3% | 84.1–97.4 | Passed 56/60 | mlx_lmextraction | 2026-09-08 |
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.
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-mlx-mixed_3_6/ (3 shards) |
| Format | MLX --quant-predicate mixed_3_6 on a 4-bit g64 base, 14.94 GB of shards |
| Where from | made here 2026-09-08 from the bf16 |
| Status | deleted 2026-09-16 20:10Z |
| Hugging Face | never uploaded |
How it was made
- Input:
models/glm-4.7-flash-bf16(GLM-4.7-Flash bf16 safetensors (zai-org)). - Command:
B=mlx-main-venv/bin # mlx-lm 0.32.0, git main @ 7fb4be44; mlx core 0.32.2
$B/mlx_lm.convert --hf-path models/glm-4.7-flash-bf16 \
--mlx-path models/glm-4.7-flash-mlx-mixed_3_6 \
-q --q-bits 4 --q-group-size 64 --quant-predicate mixed_3_6mixed_3_6 is mlx_lm's built-in predicate. It keeps the 4-bit g64 base and moves sensitive layers to other widths.
- Date: 2026-09-08, between the 07:51Z queueing (behind the Laguna GGUF download) and the 10:52Z screening run; exact build time not on record. Who asked: Alan, "We should try mixed precision etc so keep the bf16 GLM".
- Bits per weight: not on record (the build log line was captured only for mixed_4_6).
- Size: 3 safetensors shards, 5.37 + 5.35 + 4.22 = 14.94 GB; du 14 G; "13.9 GB".
- Metadata trap:
config.jsonreports "4-bit g64 affine", the same as plain q4. "Only the file sizes and the build log give it away".
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
glm-4.7-flash-mlx-mixed_3_6 | 2026-09-08 ~10:5xZ | mlx-main-venv/bin/python -m mlx_lm server (mlx-lm 0.32.0; run log mlx_lm 0.32.0) | defaults | extraction | screening only, "the only cross-comparable suite on MLX" |
No seats held.
Decisions
- Best size/quality of the four MLX builds: perplexity 11.677 at 14 G, within 0.53 of q8 at under half the size. Became the upload candidate with q8 ("publish
q8andmixed_3_6"). It was the recipe reused for granite-guardian-4.1-8b-mlx-mixed36 ("mixed 3/6 beat q4 on GLM perplexity at smaller size"). - Survived the 2026-09-09 sweep as the "HF upload candidate".
- Deleted 2026-09-16 20:10Z as SUPERSEDED: "slower than the Q4_0 beside it (byte-alignment-governs-decode)". Alan: "Does that leave another GLM flash", then "Go for it then". (judgement worth review) No speed was measured for this build; the reason is inferred from a finding. It was also never uploaded, though it was the upload candidate.
What went wrong
- Perplexity first failed on all four builds:
mlx_lm.perplexityneedsdatasetswithout declaring it, and the next attempt hit an unknown split. The third run, on--data-path eval/ppl-corpus, worked. - The screening result (93.3, behind q4's 96.7) pointed the wrong way, and at n=60 could not separate the builds. Perplexity reversed it.
What happened to it
Made here 2026-09-08 → screening + perplexity 2026-09-08 → deleted 2026-09-16 20:10Z. No RESTORE entry; never on HF. Re-creatable with the command above, given the bf16 and the same mlx-lm commit. The predicate's layer choices may change between mlx-lm versions.