GLM-4.7-Flash · file← GLM-4.7-Flash
MLX 8-bit g64
Our 8-bit build for Apple's MLX. Slower than llama.cpp, and never uploaded. Made here from bf16.
- Size
- not recorded
- Status
- Deleted 9 Sep 2026
- Scored runs
- 3
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| groundingglm-4.7-flash-mlx-q8unenforced | 76.8% | 68.7–83.3 | Passed 96/125 | mlx_lmextraction | 2026-09-08 |
| nullschemaglm-4.7-flash-mlx-q8unenforced | 59.7% | 48.2–70.3 | Passed 43/72 | mlx_lmextraction | 2026-09-08 |
| screeningglm-4.7-flash-mlx-q8 | 96.7% | 88.6–99.1 | Passed 58/60 | 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-q8/ |
| Format | MLX 8-bit g64 affine, du 30 G |
| Where from | made here 2026-09-08 from the bf16 |
| Status | deleted 2026-09-09 14:08Z |
| Hugging Face | never uploaded |
How it was made
- Input:
models/glm-4.7-flash-bf16(GLM-4.7-Flash bf16 safetensors (zai-org)), so a single quantisation pass from bf16. - Command:
PY=mlx-main-venv/bin/python # mlx-lm 0.32.0, git main @ 7fb4be44d560e5b74595210f83cb6003a57e52a7; mlx core 0.32.2
$PY -m mlx_lm convert --hf-path models/glm-4.7-flash-bf16 \
--mlx-path models/glm-4.7-flash-mlx-q8 -q --q-bits 8 --q-group-size 64Same command in two one-off scripts (the second is the one whose log shows the conversion, "=== convert GLM q8 00:38 ===", "=== convert GLM q4 00:39 ===" local). The conversion's own log lines are not on record.
- Date: 2026-09-08, ~00:3x BST (23:3xZ on 2026-09-07). Who asked: Alan, "Can we quantize it?" and "First model for my hugging face profile :)".
- Size: du 30 G. Exact bytes not on record; never uploaded.
config.jsonreports8-bit g64 affine.
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-q8 | 2026-09-07 23:3xZ–23:5xZ | mlx-venv/bin/python -m mlx_lm server --model <dir> (bench_mlx); run log: mlx_lm 0.31.3 (see below) | mlx_lm server defaults; no flags | extraction | the first MLX build of the model, on the three extraction suites, beside the GGUF Q8_0 baseline |
No seats held.
Decisions
- Made as the bit-matched counterpart of the GGUF Q8_0: "q8 first: matches the existing GGUF q8 baseline for a like-for-like runtime comparison".
- Upload candidate. A card was drafted for
freshcontext/GLM-4.7-Flash-MLX-8bit. After perplexity the plan was "publishq8andmixed_3_6, dropq4". Never uploaded; no decision on record. - Deleted 2026-09-09 14:08Z, "mixed_3_6 covers it". Alan: "ok remove the granite extras and the superseded models". (judgement worth review) This was the perplexity reference (11.149) and the upload's lead build.
What went wrong
- Version stamp in doubt. The run log says
mlx_lm 0.31.3, the server venv's pinned version (mlx-venv). The card drafted the same night saysglm4_moe_liteis not in 0.31.3 and needs git main. One of the two is wrong;mlx-venvno longer exists to check.
What happened to it
Made here 2026-09-08 → benched 2026-09-08 → deleted 2026-09-09 14:08Z. No RESTORE entry; never on HF. Re-creatable with the command above, given the bf16 and mlx-lm at or after 7fb4be44. A later mlx-lm may produce different bytes.