FreshContextA context lab · Fife
Drawing
02 · Benchmarks
Sheet
2 of 3
Rev
A · 2026-10-01

Qwen3.5-4B · file← Qwen3.5-4B

MLX 4-bit + MTP head

An MLX build with an extra head for drafting tokens ahead, for a speed test. Still on disk, in a folder our clear-outs never looked in. MTPLX Optimized-Speed, Youssofal.

Size
2.4 GB
Status
On disk
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.mtplx/models/Youssofal--Qwen3.5-4B-MTPLX-Optimized-Speed/ (model.safetensors 2,367,238,661 B + mtp.safetensors 86,701,062 B)
FormatMLX affine 4-bit g64 body + BF16 MTP head
Where fromdownloaded 2026-08-16 ~13:35Z (file mtimes, local 14:35–14:43) via MTPLX's model cache
Statusstill on disk, outside models and outside every sweep
Hugging Facehttps://huggingface.co/Youssofal/Qwen3.5-4B-MTPLX-Optimized-Speed

Where it came from

  • Repo: Youssofal/Qwen3.5-4B-MTPLX-Optimized-Speed. HF cache snapshot 5a8d32823f26573b40bdf027844647d6461a4fc8 holds only config.json and mtplx_runtime.json (.cache/huggingface/hub/models--Youssofal--…, 16 KB).
  • Held at: .mtplx/models/Youssofal--Qwen3.5-4B-MTPLX-Optimized-Speed/ — model.safetensors 2,367,238,661 B, mtp.safetensors 86,701,062 B, plus tokenizer/processor files; 2.4 GB total (ls -la, du -sh, 2026-10-23). sha256 not on record.
  • Date: 2026-08-16, file mtimes 14:35–14:43 local (13:35–13:43Z). The fetch command is not on record (no transcript covers 2026-08-16); the path pattern is MTPLX's own model cache.
  • What it is (its config.json / mtplx_runtime.json, read 2026-10-23): base trunk Qwen/Qwen3.5-4B; body MLX affine 4-bit group 64; MTP head kept BF16 (mtp_policy keep_bf16); "forged" by the uploader 2026-07-18 with MTPLX's forge; arch_id qwen3-next-mtp. README: "4-bit (group 64) Qwen3.5-4B with a calibrated native-MTP draft head … 2.47 GB on disk", M5 Max claim 133.6 → 227.8 tok/s (1.71×).
  • Who asked / why: not on record. The A/B script's docstring states the question: whether MTPLX's published 1.6× (M4 mini) / 2.24× (M5 Max) holds on an M1 Max that reports hardware_acceleration_eligible: false (eval/mtplx_ab.py).

How it was served

labeldates (UTC)binary and versionctx / -np / KV / -fa / --jinja / --reasoning-formatharness profilewhy this label exists
— (no run-log label; MTPLX model id qwen3.5-4b-mtplx-optimized-speed)2026-08-16 (≈13:44–14:24Z, from script and result-file mtimes)mtplx/.venv/bin/mtplx daemon; version on that date not on record (v2.7.1 was installed by 2026-09-11)MTPLX-managed; depth 0–3 set live via mtplx settings depth=<d>none (own scripts: 400-token chat/code/math prompts, temperature 0; then 6 interleaved AR/MTP pairs per depth on one coding prompt, MAX_TOK 9000)the MTP speed A/B (eval/mtplx_ab.py, eval/mtplx_paired.py)

Seats held: none.

Decisions

  • Downloaded for the MTP A/B (reconstructed from the scripts).
  • 2026-09-11 — not re-tested. When Alan said "the other runtime we haven't tested yet is mtplx", the assistant found the August result and argued a 4B is the wrong subject: "A 4B target at Q4 reads ~2.4 GB per token — far too cheap to speculate for profitably". The thread then moved to larger models; MTPLX was not re-run on this build.
  • Never benched for quality, never seated, never published (it is a third-party upload).

What went wrong

  • Void first A/B. results_mtplx_ab.json recorded zeros for every arm; the paired script replaced it within the hour (script mtimes 14:44 and 15:24 local). Cause not on record.
  • Outside every sweep. The file lives in .mtplx, which the 2026-10-11 sweeps (scoped to models and the HF cache) never listed. It is the only Qwen3.5-4B weights still on the box.
  • Possibly a pre-fix sidecar. The README's "July 2026 rebuild" section says earlier revisions "shipped a defective MTP sidecar … MTP made the model slower than plain decoding", and that MTPLX 2.2.0+ heals it at load. The held forge date (2026-07-18) and the August MTPLX version cannot settle whether the SLOWER verdict was measured on a defective head.

What happened to it

Downloaded 2026-08-16 → MTP A/B 2026-08-16 (SLOWER at depths 1 and 2) → never suite-benched → on disk at .mtplx/models/Youssofal--Qwen3.5-4B-MTPLX-Optimized-Speed/ (2.4 GB, checked 2026-10-23). No RESTORE entry (never deleted). Re-downloadable from Youssofal/Qwen3.5-4B-MTPLX-Optimized-Speed (revision may have changed).