BTL-4 · file← BTL-4
Q4_K_M
The one we kept: a community conversion, about twice as fast in our runs and better with tools. Still on disk, set up to swap in. Community DogukanUrker GGUF; stock b10893.
- Size
- 21.5 GB
- Status
- On disk
- Scored runs
- 15
- Smallest Mac
- 36 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| chainwalkbtl-4-q4km | 0.0% | 0.0–65.8 | Passed 0/2 | llama.cppagentic-fast | 2026-09-15 |
| contractbtl-4-q4km | 70.8% | 56.8–81.8 | Passed 34/48 | llama.cppextraction | 2026-09-15 |
| enrichmentbtl-4-q4km | 81.5% | 69.2–89.6 | Passed 44/54 | llama.cppextraction | 2026-09-15 |
| epistemicsbtl-4-q4km | 100.0% | 92.6–100.0 | Passed 48/48 | llama.cppstudy-nothink | 2026-09-15 |
| groundingbtl-4-q4km | 76.8% | 68.7–83.3 | Passed 96/125 | llama.cppextraction | 2026-09-15 |
| nullschemabtl-4-q4km | 44.4% | 33.5–55.9 | Passed 32/72 | llama.cppextraction | 2026-09-15 |
| screeningbtl-4-q4km | 91.7% | 81.9–96.4 | Passed 55/60 | llama.cppextraction | 2026-09-15 |
| screeningbtl-4-q4km-recheck | 91.7% | 81.9–96.4 | Passed 55/60 | llama.cppextraction | 2026-09-17 |
| synthesisbtl-4-q4km | 83.3% | 72.0–90.7 | Passed 50/60 | llama.cppextraction | 2026-09-15 |
| taggerbtl-4-q4km | 87.1% | 82.2–90.7 | Passed 209/240 | llama.cppextraction | 2026-09-15 |
| tagger_nullablebtl-4-q4km68.3% coverage | 71.3% | 64.0–77.7 | Passed 117/164 | llama.cppextraction | 2026-09-15 |
| toolcallbtl-4-q4km | 93.3% | 84.1–97.4 | Passed 56/60 | llama.cppextraction | 2026-09-15 |
| toolsbtl-4-q4km | 100.0% | 79.6–100.0 | Passed 15/15 | llama.cppagentic-fast | 2026-09-15 |
| toolscalebtl-4-q4km | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppagentic-fast | 2026-09-15 |
| worldviewbtl-4-q4km | 55.0% | 39.8–69.3 | Passed 22/40 | 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.
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-Q4_K_M.gguf (21,544,204,064 B) |
| Format | Q4_K_M (community) |
| Where from | downloaded 2026-09-15 |
| Status | benched; swap-seat bench entry btl-4 in fleetview; on disk |
| Hugging Face | DogukanUrker/BTL-4-GGUF |
Where it came from
- Source:
DogukanUrker/BTL-4-GGUF, fileBTL-4-Q4_K_M.gguf, fetched withhf_hub_downloadintomodels/btl-4-q4km. The card says it was converted from the BF16 source withconvert_hf_to_gguf.py --no-mtp, token embeddings and output at q8_0. Revision and sha256 not on record. - Size: 21,544,204,064 B (20.1 GiB).
- Dates: download started 2026-09-15 07:51Z; ready by 08:11Z.
- Who asked: Alan, 07:50Z, "I think I'd rather test the q4_0 first then 8b"; no Q4_0 exists, so the assistant fetched the only other byte-friendly build. The community card also reported 52.0-52.4 tok/s on an RTX 3060, which the assistant used as independent support for the quant diagnosis.
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-q4km | 2026-09-15 08:17-09:16Z (run log local 09:17:28-10:16:43) | stock .local/bin/llama-server, b10893 (c32d1dabe), manifest b10893-c32d1dabe | -c 32768 (8192 per slot) / 4 / q4_0 / on / on / deepseek (--cache-ram 2048, -ngl 999, --alias btl-4-q4km) | extraction; agentic-fast for tools, toolscale, chainwalk; study-nothink for epistemics | the 14-suite board on the byte-aligned build after the IQ2 run |
btl-4-q4km-recheck | 2026-09-17 17:36Z (run log local 18:36:41) | llama.cpp/build/bin/llama-server (b10893 lineage; manifest b10893-c32d1dabe) | -c 8192 / 1 / default / on / on / not set (--no-warmup, -ngl 99) | extraction, screening only | spot-check that a recorded score reproduces on the weights still on disk, for the provenance-stamp finding |
| — (no label; speed) | 2026-09-15 ~08:20Z | same as btl-4-q4km | same | none | three speed requests: 48.5 t/s decode, 84 t/s prefill on run 3; 20.5 GB resident |
| — (no label; seat race probes) | 2026-09-18 20:40-22:00Z | .local/bin/llama-server | not on record | none | contradiction/1, edgefill/1 probes and a same-prompt seat-speed table: 666 t/s prefill, 49.9 t/s decode, 21.5 GB weights, 80 KB/token KV |
Seats held: none permanently. fleetview has a swap-seat bench entry btl-4 (.local/bin/llama-server, -c 65536 -np 1 --alias btl-4, label "BTL-4 Q4_K_M", note "generalist · 666 pf / 50 dec"), added in the 2026-09-20 hot-swap work. It was in a swap-seat contest on 2026-09-18/19 against Ornith and bonsai-2 and lost.
Decisions
- Downloaded after the IQ2 board showed a quant-bound model, to give BTL-4 a fair test.
- Not seated (2026-09-18): best generalist but 20 GB and worst of the big models on worldview.
- Kept through the 2026-09-16 inventory ("tested, no faster build on disk") and left out of the 2026-10-11 sweeps; still on disk; present as a swap-seat bench entry.
- Published as
BTL-4 (Q4_K_M)with status candidate; the write-up's verdict quoted mean scores as percentages.
What went wrong
- Server killed mid-suite (reconstructed). At 09:16Z the assistant read "q4km suites: 14/14" (a count of suite headers started, not finished), printed the board, and killed the server in the same turn. The last 76 tagger_nullable cases got connection-refused. The cause is reconstructed from the error pattern and timing; the run log's two rows (10:16:43 and 10:46:43 local) then disagree on coverage.
- Binary/flag confound with the IQ2 run: see BTL-4 IQ2_XXS (vendor BTL-4-Compact download; prism-fork run) incidents.
- Missing flags (standing check): this build has
--jinjaand--reasoning-format deepseek; the recheck lacks--reasoning-formatand uses-np 1, and still matched 91.7% -> 91.7% on screening. - Stamp check: the manifest build
b10893-c32d1dabematches the stock binary named in the launch script, so no mis-stamp here.
What happened to it
Downloaded 2026-09-15 07:51Z -> speed + 14-suite board -> server killed 09:16Z -> recheck 2026-09-17 -> seat-race probes 2026-09-18 -> bench entry in fleetview 2026-09-20 -> on disk 2026-10-23 (models/btl-4-q4km/BTL-4-Q4_K_M.gguf). No RESTORE entry (not deleted). Re-downloadable from DogukanUrker/BTL-4-GGUF.