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

Model← Models

Qwen3.8-9B-Distill (Q8_0)

8.9B dense, Q8_0, 9.8 GB.

Parameters
8.9B
Quant
Q8_0
Size
9.8 GB
Family
qwen
RecordStatus · runtime · seat

qwen3.8-9b-distill

Measured, not seateddenseresidentllama.cpp

Model card ↗

Outcome as of 20 September 2026, from the seat registry. The write-up below was made when the model was tested.

Why it was tested

Full-parameter distillation of Qwen3.8 2.4T A95B into the Qwen3.5-9B architecture, ~70k curated teacher traces. Tested as a cheap utility seat.

What came of it

A single-turn processor, not a driver. Strong where the work is 'judge this thing I handed you': epistemics 95.8, enrichment 95.9, screening 91.7, tagger 94.4/96.3 at n=240, and synthesis 83.3 matching its 27B teacher exactly. Collapses where the work is deciding: toolcall 0/60, contract 25.0, nullschema 18.1. 30.8 tok/s in 12 GB at 64k.

Recorded runs14 scored · 95% intervals
SuiteScore95% interval PassedRuntime · profileRecorded
chainwalk0.0%0.0–65.8Passed 0/2llama.cppagentic-fast2026-09-14
contract25.0%14.9–38.8Passed 12/48llama.cppextraction2026-09-14
enrichment79.6%67.1–88.2Passed 43/54llama.cppextraction2026-09-14
epistemics89.6%77.8–95.5Passed 43/48llama.cppstudy-nothink2026-09-14
grounding67.2%58.6–74.8Passed 84/125llama.cppextraction2026-09-14
nullschema18.1%10.9–28.5Passed 13/72llama.cppextraction2026-09-14
screening91.7%81.9–96.4Passed 55/60llama.cppextraction2026-09-14
synthesis83.3%72.0–90.7Passed 50/60llama.cppextraction2026-09-14
tagger83.3%78.1–87.5Passed 200/240llama.cppextraction2026-09-14
tagger_nullable90.0%85.6–93.2Passed 216/240llama.cppextraction2026-09-14
toolcall0.0%-0.0–6.0Passed 0/60llama.cppextraction2026-09-14
tools100.0%79.6–100.0Passed 15/15llama.cppagentic-fast2026-09-14
toolscale100.0%91.2–100.0Passed 40/40llama.cppagentic-fast2026-09-14
worldview40.0%26.3–55.4Passed 16/40llama.cppextraction2026-09-14

Runs marked unenforced were measured through mlx_lm on a suite that requires constrained decoding. mlx_lm accepts the schema and ignores it, so those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.

Every run log entry for this model (14)
FinishedSuiteRuntime · profileScoreCoverageGrader · dataset · config
tagger_nullabletagger-nullable/1not recordedextraction90.0%Coverage 240/240g 04e875431e26c4b5d db3addf23a2bc4cec d7923863e8b56e76
taggertagger/2not recordedextraction83.3%Coverage 240/240g 63e04d139a0e5ed3d 7ef587bf33e5788ac d2642a91c9a03367
enrichmentenrichment/3not recordedextraction79.6%Coverage 54/54g b51222c2ca38cfdfd 75361864baa727abc d537736901f01a81
contractcontract/2not recordedextraction25.0%Coverage 48/48g cc2d23c5bae45474d 1fe53d30cb5480f6c 5a052624057c52ff
chainwalkchainwalk/3not recordedagentic-fast0.0%Coverage 2/2g f4bfcf6f040b0978d 129a9e74408c8919c 43d9a11df6da4927
epistemicsepistemics/4not recordedstudy-nothink89.6%Coverage 48/48g acf84972a0ad5f38d 4bce453fd376553cc a0e9adc5c6542d53
groundinggrounding/1not recordedextraction67.2%Coverage 125/125g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 696fd8b30eed53f0
screeningscreening/1not recordedextraction91.7%Coverage 60/60g 0661b49a10be2d2ad 64c97ca2269c29afc 5ac25f36beb78f8c
synthesissynthesis/2not recordedextraction83.3%Coverage 60/60g 4156dd3c13361e8bd 267d1954e225c5b6c 8481e2a6e1f3e60c
worldviewworldview/2not recordedextraction40.0%Coverage 40/40g 9acfba1b704feb28d e97835de9e6fc227c 6a67c959e17c98f5
nullschemanullschema/1not recordedextraction18.1%Coverage 72/72g d7448e0f5fc3f9afd 3b91280bb40372efc 56c7dc8f982d59d3
toolstools/1not recordedagentic-fast100.0%Coverage 15/15g 2763675a5c3da564d 33a44812f0ad4c39c 2411d23420a9fcb7
toolscaletoolscale/1not recordedagentic-fast100.0%Coverage 40/40g a2bc3751abfa28e0d b24abb108c409178c 13a2f2f87c9bcdc8
toolcalltoolcall/1not recordedextraction0.0%Coverage 60/60g 804e8fc9457649d8d 10af88c170513155c 2cd02d2d291d660a