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

Qwen3.6-35B-A3B · file ← Qwen3.6-35B-A3B

UD-Q5_K_M with MTP head

The first copy and the everyday model from 11 August. All the llama.cpp tests ran on it; still set up to load on demand. Unsloth GGUF.

Size
27.1 GB
Status
On disk
Scored runs
8
Smallest Mac
48 GB estimate, 8K context
Results 8 scored · 95% intervals
SuiteScore95% interval PassedRuntime · profileRecorded
chainwalk qwen3.6-35b-ag void50.0%9.5–90.5Passed 1/2 llama.cpp extraction2026-09-07
grounding qwen3.6-35b82.4%74.8–88.1Passed 103/125 llama.cpp extraction2026-08-27
nullschema qwen3.6-35b61.1%49.6–71.5Passed 44/72 llama.cpp extraction2026-08-27
screening qwen3.6-35b86.7%75.8–93.1Passed 52/60 llama.cpp extraction2026-08-27
tagger qwen3.6-35b-tag90.8%86.5–93.9Passed 218/240 llama.cpp extraction2026-08-31
tagger_nullable qwen3.6-35b-tag92.5%88.5–95.2Passed 222/240 llama.cpp extraction2026-08-31
tools qwen3.6-35b-ag93.3%70.2–98.8Passed 14/15 llama.cpp extraction2026-09-07
toolscale qwen3.6-35b-ag80.0%65.2–89.5Passed 32/40 llama.cpp extraction2026-09-07

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.

Corrections 1 to the record published 21 October 2026
  • On the chainwalk run of 7 September 2026: voided. This row is a duplicate of the 20:19:37 row, made by re-scanning the same result file, so it is voided and the later row is the published one.

    Corrected 26 October 2026 · a mechanical fix

The record How 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.

Filemodels/qwen3.6-35b-a3b/Qwen3.6-35B-A3B-UD-Q5_K_M.gguf
FormatGGUF UD-Q5_K_M (Unsloth imatrix), 27,087,812,896 B
Where fromdownloaded 2026-08-11 from unsloth/Qwen3.6-35B-A3B-MTP-GGUF, sha256 verified
Statuson disk; router preset [qwen3.6-35b] (on demand, not a fleetview seat); benched
Hugging Facehttps://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF

Where it came from

  • Repo / files: unsloth/Qwen3.6-35B-A3B-MTP-GGUF, Qwen3.6-35B-A3B-UD-Q5_K_M.gguf and mmproj-F16.gguf (models/qwen3.6-35b-a3b/finish.sh, supervise.sh, read 2026-10-23).
  • Commands: supervise.sh — curl -sL -C - --connect-timeout 20 --speed-limit 200000 --speed-time 20 -o "$f" "$B/$f" with up to 500 retries; the main file finished through a 6-connection range downloader (pget.log: "done: 27087MB / 27087MB"); finish.sh fetched the mmproj with 4 connections and checked both against the HF tree API LFS oids.
  • Date: 2026-08-11, complete 15:03 BST (14:03Z) (file mtime; .dl_complete 15:05 BST).
  • Size: 27,087,812,896 B (finish.log; stat 2026-10-23). mmproj 899,283,584 B.
  • sha256: 8ca8224c5519b50f286b9739f9ab1c2bc0759e02cfb6cbf48a09f0759eb2d319 — "OK" against the HF LFS oid (expected.txt, finish.log). mmproj 71f3cbc1f7cc0f30d09d41cfa924c0060827ebc33bf15ace7e86661e856f0160 OK.
  • Revision: not on record (resolved main).
  • Quant recipe (GGUF header): general.quantized_by = Unsloth, file_type 17 (Q5_K_M, "UD" dynamic), imatrix imatrix_unsloth.gguf over unsloth_calibration_Qwen3.6-35B-A3B.txt, 510 entries, 77 chunks; nextn_predict_layers = 1 (the MTP head; 4 nextn tensors).
  • Smoke test: llama-cli on build b10369-6e62ba538, 2026-08-11 15:18 BST, "Prompt: 120.4 t/s | Generation: 42.3 t/s" (smoke.log).
  • Who / why: chosen over the non-MTP build for the embedded head (a lab note). Alan's request is not on record.

How it was served

labeldates (UTC)port / binaryflags that matteredharness profilewhy this label exists
(none — interactive)2026-08-11 → ~2026-08-27router, llama.cpp/build/bin/llama-server (b10369 until the 2026-08-25 rebuild to b10618)preset [qwen3.6-35b]: mmproj, spec-type = draft-mtp, c = 131072; globals n-gpu-layers = 99, jinja = true; launch --models-preset fleet-models.ini --models-max 1 --context-shift—daily driver / DSH model (a lab note)
qwen3.6-35b2026-08-27 07:00Z–07:04Zrouter, b10618 (eb25b7263)same preset; -np not set (server default, not on record); no --reasoning-formatextraction (thinking off, max_tokens 2048)first locked-suite runs: grounding, screening, nullschema
qwen3.6-35b-tag2026-08-31 ~17:00Z–17:30Zstandalone server, b10618--ctx-size 16384 --n-gpu-layers 99 --jinja; no MTP, no mmprojextractiontagger head-to-head with gemma4-12b and Qwen3-Next under identical standalone flags
qwen3.6-35b-ag2026-09-07 ~18:13Z–18:20Zstandalone server, b10618--ctx-size 32768 --n-gpu-layers 99 --jinja; no MTP/mmproj; tool-call gate at 256 tokensextractionagentic suites (tools, toolscale, chainwalk) with Qwen3-Next and AgentWorld

Run-log runtime for all three labels: llama.cpp b10618 (eb25b7263) (runtime_source "evidence"), matching the router and standalone binary. The qwen3.6-35b label also carries one toolcall row that is not this file (Qwen3.6-35B-A3B MLX 4-bit (mlx-community)).

Seats held: default router model and DSH driver from 2026-08-11 (fleet-models.ini "DEFAULT — best balance"); the load-on-startup setting removed 2026-08-20; still a router preset on 2026-10-23 (fleet-models.ini). Never a fleetview seat.

Decisions

  • Downloaded and made default 2026-08-11 (family decision log). Kept on disk through every sweep; it has run-log results and remained a router preset.
  • Not converted to Q4_0 despite the 2026-09-11 audit flag (no decision on record).
  • Not published to HF (a third-party quant).

What went wrong

  • Never benched for 16 days. In the router but absent from the run log until 2026-08-27, because it entered as an interactive model.
  • Partial row recorded as a result. scan(record=True) ran mid-suite and logged grounding 86.3 on 73 cases; the complete 82.4 (103/125) was appended after. The run log still shows both, the first unmarked.
  • Smoke-gate false alarm. The one-tool-call gate with max_tokens=16 reported empty content for this model during the tagger run because reasoning ate the budget.
  • Harness thinking defaults (2026-08-22). Three board suites failed because the template thinks by default and the harness never sent enable_thinking: false.
  • "Exactly one copy" misread on 2026-09-12: the assistant said this 27 GB file was the only copy, missing the 19 GB MLX build in the HF cache.

What happened to it

Downloaded 2026-08-11 → default/DSH driver 2026-08-11 → benched 2026-08-27, 08-31, 09-07 → displaced as interactive model by the MLX build 2026-09-08 → router preset only → on disk 2026-10-23 (27,087,812,896 B). No RESTORE-deleted-*.sh entry needed. Re-downloadable from unsloth/Qwen3.6-35B-A3B-MTP-GGUF (verify against the sha256 above).