Qwen3.6-35B-A3B · file← Qwen3.6-35B-A3B
Splash package
A package for the Splash runtime: the fastest this model ran on short requests, but below the quality bar. Incoai, q4 MoE.
- Size
- 19.5 GB
- Status
- On disk
- Scored runs
- 7
- Smallest Mac
- 36 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| epistemicsqwen3.6-a3b-splash-nothink | 100.0% | 92.6–100.0 | Passed 48/48 | splashstudy-nothink | 2026-09-28 |
| groundingqwen3.6-a3b-splash-nothink | 70.4% | 61.9–77.7 | Passed 88/125 | splashstudy-nothink | 2026-09-28 |
| nullschemaqwen3.6-a3b-splash-nothink | 34.7% | 24.8–46.2 | Passed 25/72 | splashstudy-nothink | 2026-09-28 |
| synthesisqwen3.6-a3b-splash-nothink | 78.3% | 66.4–86.9 | Passed 47/60 | splashstudy-nothink | 2026-09-28 |
| toolcallqwen3.6-a3b-splash-nothink | 56.7% | 44.1–68.4 | Passed 34/60 | splashstudy-nothink | 2026-09-28 |
| toolscaleqwen3.6-a3b-splash-nothink | 77.5% | 62.5–87.7 | Passed 31/40 | splashstudy-nothink | 2026-09-28 |
| worldviewqwen3.6-a3b-splash-nothink | 40.0% | 26.3–55.4 | Passed 16/40 | splashstudy-nothink | 2026-09-28 |
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 | HF cache models--incoai--Qwen3.6-35B-A3B-Splash @ 0f4714b2 |
| Format | Splash package, splash-packed-q4-moe (q4/q8, group 64), target 19,541,082,112 B + DFlash2 draft 479,444,992 B + vision 901,939,200 B = 20,922,466,304 B |
| Where from | downloaded 2026-09-28 by Splash make install |
| Status | on disk; benched; "benchmark datapoint only" |
| Hugging Face | https://huggingface.co/incoai/Qwen3.6-35B-A3B-Splash |
Where it came from
- Repo:
incoai/Qwen3.6-35B-A3B-Splash, snapshot0f4714b2db37b5f3c42a10de07281e74f88e4adc(HF cache, read 2026-10-23). - Command:
HF_XET_HIGH_PERFORMANCE=1 make install MODEL=incoai/Qwen3.6-35B-A3B-Splashinbuilds/splash-m1. - Size: 20,922,466,304 B: target 19,541,082,112 + DFlash2 draft 479,444,992 + vision 901,939,200. sha256: not on record; whether
make installhash-verified this package is not on record (a lab note says so for the dense package). - Format: Splash schema 4 / splash-packed-q4-moe (q4/q8, group 64), with its own trained DFlash 2 draft.
- Who / why: Alan, after the dense package: "I suppose we have to do the qwen 3.6 too? that's done btw", having said earlier "I'm less interested in 3.6 as I have better models". The assistant's reason: same A3B class as the ornith judgment seat, so it shows what Splash's MoE path could do for that class.
How it was served
| label | dates (UTC) | port / runtime | flags that mattered | harness profile | why this label exists |
|---|---|---|---|---|---|
| (speed only, unlabelled) | 2026-09-28 10:43Z–10:45Z | Splash M1 fork ghaslbe/splash-apple7-m1-kernels-fork @ 6d45618 | ./splash serve --model incoai/Qwen3.6-35B-A3B-Splash --max-memory 28G | depth-swept SSE protocol, not the harness | speed arm vs ornith |
qwen3.6-a3b-splash-nothink | 2026-09-28 15:09Z–15:15Z | same | --default-reasoning-effort none (Splash ignores chat_template_kwargs.enable_thinking) | study-nothink (temperature 0, seed 42, max_tokens 10000) | quality battery, seven suites |
Harness command: python3.12 -m harness.runner --suite $s --model <server> --model-id incoai/Qwen3.6-35B-A3B-Splash --label qwen3.6-a3b-splash-nothink --profile study-nothink. Run-log runtime splash splash-m1 6d45618 matches git -C builds/splash-m1 log -1 (2026-10-23). Schema enforcement: sentinel probe passed, 0/97 constrained responses violate.
Seats held: none.
Decisions
- No roster role: "every roster role is quality-gated and it fails the gates its speed would serve", after Alan's "so no place for the a3b then... speeds were juicy".
- Kept in the HF cache (not swept on 2026-10-11).
What went wrong
- The thinking leak that voided the first dense battery (
qwen3.8-27b-splash) was avoided here by serving with--default-reasoning-effort nonefrom the start. - An earlier lab note states that "no same-model llama.cpp comparator for Qwen3.6 exists on the current harness". The UD-Q5_K_M runs on llama.cpp exist (grounding 82.4, nullschema 61.1, screening 86.7,
extractionprofile).
What happened to it
Downloaded 2026-09-28 → speed arm and quality battery 2026-09-28 → not seated → on disk 2026-10-23 (HF cache, ~20 GB). No RESTORE entry needed. Re-downloadable from incoai/Qwen3.6-35B-A3B-Splash @ 0f4714b2.