BTL-4 · file← BTL-4

IQ2_XXS

Bad Theory Labs' own file and the only quantised one they published, at 2.3 bits a weight. Tested first, then set aside for the Q4_K_M; still on disk. Vendor BTL-4-Compact download; prism-fork run.

Size: 10.0 GB
Status: On disk
Scored runs: 14
Smallest Mac: 18 GB estimate, 8K context

Results14 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
chainwalkbtl-4	0.0%	0.0–65.8	Passed 0/2	llama.cppagentic-fast	2026-09-15
contractbtl-4	64.6%	50.4–76.6	Passed 31/48	llama.cppextraction	2026-09-15
enrichmentbtl-4	87.0%	75.6–93.6	Passed 47/54	llama.cppextraction	2026-09-15
epistemicsbtl-4	100.0%	92.6–100.0	Passed 48/48	llama.cppstudy-nothink	2026-09-15
groundingbtl-4	76.8%	68.7–83.3	Passed 96/125	llama.cppextraction	2026-09-15
nullschemabtl-4	36.1%	26.0–47.6	Passed 26/72	llama.cppextraction	2026-09-15
screeningbtl-4	95.0%	86.3–98.3	Passed 57/60	llama.cppextraction	2026-09-15
synthesisbtl-4	80.0%	68.2–88.2	Passed 48/60	llama.cppextraction	2026-09-15
taggerbtl-4	85.0%	79.9–89.0	Passed 204/240	llama.cppextraction	2026-09-15
tagger_nullablebtl-4	91.7%	87.5–94.5	Passed 220/240	llama.cppextraction	2026-09-15
toolcallbtl-4	76.7%	64.6–85.6	Passed 46/60	llama.cppextraction	2026-09-15
toolsbtl-4	100.0%	79.6–100.0	Passed 15/15	llama.cppagentic-fast	2026-09-15
toolscalebtl-4	100.0%	91.2–100.0	Passed 40/40	llama.cppagentic-fast	2026-09-15
worldviewbtl-4	62.5%	47.0–75.8	Passed 25/40	llama.cppextraction	2026-09-15

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.

Corrections16 to the record published 21 October 2026

Build of llama.cpp on 14 runs was 0.4.0-dev (build 10893, commit c32d1dabe), now 0.2.0-dev (build 1, commit 6ac5eb0). Served by the PrismML llama.cpp fork (build 1, commit 6ac5eb0), as every run manifest records; the stamp named the stock binary instead.
Corrected 24 October 2026 · a mechanical fix
Made from was btl-4-q4km, now none. The IQ2_XXS is the vendor's own build and the Q4_K_M a third-party conversion; neither was made from the other.
Corrected 24 October 2026 · a mechanical fix
Verdict corrected. Worldview moved the other way, and the two builds ran on different runtimes and flags.
What it said before
Verdict: Rejected in favour of the Q4_K_M build. Same weights, and the quantisation alone cost 27 points of worldview and 17 of toolcall, plus 2.06x throughput (23.6 vs 48.5 tok/s). Evidence for iq-codebook-penalty-transfers-to-moe-kquant-does-not.
Corrected 24 October 2026 · a headline fix

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	`BTL-4-IQ2_XXS.gguf` (9,967,966,240 B)
Format	IQ2_XXS, 2.30 bpw (header `file_type`=15 is wrong)
Where from	downloaded 2026-09-14
Status	benched, rejected; on disk (`models/btl-4-compact`)
Hugging Face	`badtheorylabs/BTL-4-Compact`

Where it came from

Source: badtheorylabs/BTL-4-Compact, file BTL-4-IQ2_XXS.gguf, the only quant the vendor publishes ("No Q4_0 format is available in this release"). Downloaded with huggingface_hub.hf_hub_download into models/btl-4-compact. Revision not on record; sha256 not on record.
Size: 9,967,966,240 B (stat), 2.30 bpw. GGUF metadata trap: general.file_type=15 (Q4_K_M) is wrong for this file; the quant comes from the filename and repo.
Dates: started 2026-09-14 21:38Z; the file's mtime is 22:46 local = 21:46Z (stat).
Who asked: Alan, 21:35Z, "Fuck it why not, it's the only a2b model and could be a driver if it's any good".

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / --jinja / --reasoning-format	harness profile	why this label exists
`btl-4`	2026-09-15 06:44-08:14Z (run log local 07:44:23-09:14:38; series script started 06:41Z)	`llama.cpp-prism/build/bin/llama-server`, PrismML fork; manifest build `b1-6ac5eb0` (the run log carried both `6ac5eb0` and `0.4.0-dev (build 10893, commit c32d1dabe)`, which is the stock binary — a mixed stamp)	`-c 32768` (8192 per slot) / 4 / q4_0 / on / on / not set (`--cache-ram 1024`, `-ngl 999`, `--alias btl-4`)	`extraction`; `agentic-fast` for tools, toolscale, chainwalk; `study-nothink` for epistemics (manifests)	the vendor build on the 14-suite board, run in series with the bonsai builds
— (no label; speed)	2026-09-15 ~06:45Z	same	same	none	three requests for the numbers 1 to 300: 23.6 t/s decode, 38-71 t/s prefill

Seats held: none. The prism binary was used because the whole series script (ternary and 1-bit bonsais) served every model with it; why the stock binary was not used for this file is not on record. At 07:51Z the assistant checked whether the stock build supports the qwen35moe architecture and concluded it does, because the IQ2 file already loads and runs.

Decisions

Tested "to see what an aggressive IQ codebook costs on a sparse MoE" (btl-4 write-up).
Rejected in favour of the Q4_K_M because of decode speed and the capability drops that the first comparison seemed to show; the corrected comparison shows a mixed picture.
Kept on disk after the Sep 16 inventory marked it KEEP ("tested, no faster build on disk") and through the 2026-10-11 sweeps, where it was wrongly called never benched (a lab note).
Published: yes, as BTL-4 Compact (IQ2_XXS), status rejected, with the incorrect "27 points of worldview" verdict.

What went wrong

Mean score published as pass rate. tagger_nullable 96.2 and enrichment 98.7 in the findings and tables were mean case scores over partial-credit checks, while the run log's pass rates are 91.7 and 87.0.
Mis-stamped build (standing check). The run log's build stamp named stock b10893 for these rows although the manifests say the prism build b1-6ac5eb0; the finding Sparsity rescues a model from K-quant misalignment and not from IQ codebook lookup said "served on llama.cpp b10893" for both builds. The speed comparison between the two quants therefore crosses binaries.
Missing flag (standing check): the card marks --reasoning-format deepseek as not optional; this file's server line has none (the Q4_K_M's does).
Concurrent load. Ornith (the Hermes driver) ran at 12.3 t/s instead of 53.4 while this gauntlet shared the GPU, and its router was found down at 07:41Z; the router outage's cause is not on record.

What happened to it

Downloaded 2026-09-14 -> speed and 14-suite board 2026-09-15 -> inventory KEEP 2026-09-16 -> still on disk 2026-10-23 (models/btl-4-compact/BTL-4-IQ2_XXS.gguf). Not deleted; no RESTORE entry. Re-downloadable from badtheorylabs/BTL-4-Compact.