Qwen3.8-27B · file← Qwen3.8-27B
UD-Q4_K_M
Used for the head-to-head with K2. Deleted once we'd made the Q4_0 from it. Unsloth GGUF, MTP head embedded.
- Size
- 16.5 GB
- Status
- Deleted 12 Sep 2026
- Scored runs
- 9
- Smallest Mac
- 32 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| nullschemaqwen38-27b-np1 | 76.4% | 65.4–84.7 | Passed 55/72 | llama.cppextraction | 2026-09-10 |
| nullschemaqwen38-rebase-b10893 | 76.4% | 65.4–84.7 | Passed 55/72 | llama.cppextraction | 2026-09-10 |
| synthesisqwen38-27b-np1 | 83.3% | 72.0–90.7 | Passed 50/60 | llama.cppextraction | 2026-09-10 |
| toolcallqwen38-27b-np1 | 100.0% | 94.0–100.0 | Passed 60/60 | llama.cppextraction | 2026-09-10 |
| toolscaleqwen38-27b-np1 | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppextraction | 2026-09-10 |
| worldviewqwen38-27b-np1 | 77.5% | 62.5–87.7 | Passed 31/40 | llama.cppextraction | 2026-09-10 |
| worldviewqwen38-rebase-b10893 | 77.5% | 62.5–87.7 | Passed 31/40 | llama.cppextraction | 2026-09-10 |
| worldviewqwen38-vendor-qwen-think | 77.5% | 62.5–87.7 | Passed 31/40 | llama.cppvendor-qwen-think | 2026-09-10 |
| PARTIAL-KILLED-nullschemaqwen38-vendor-qwen-thinkflagged | 56.2% | 33.2–76.9 | Passed 9/16 | llama.cppvendor-qwen-think | 2026-09-10 |
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 | Qwen3.8-27B-UD-Q4_K_M.gguf |
| Format | Q4_K_M (UD), 16,464,440,224 B |
| Where from | downloaded 2026-09-10 |
| Status | deleted 2026-09-12, no RESTORE entry |
| Hugging Face | same repo |
Where it came from
- Date and tool: 2026-09-10 06:55Z,
huggingface_hub.hf_hub_downloadwithHF_XET_HIGH_PERFORMANCE=1intomodels/qwen3.8-27b-gguf/. sha256 and HF revision: not on record. - Pre-download check: the assistant read the GGUF header remotely and found a complete 'blk.64' next-token-prediction layer in the main file, so the separate 1.37 GB
MTP/mtp-Qwen3.8-27B-Q4_0.ggufwas not needed. Alan had asked whether the MTP heads were present. - Who asked: Alan, 2026-09-10 06:49–06:54Z.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
qwen38-27b-np1 | 2026-09-10 07:26–07:58 | llama.cpp b10618 (eb25b7263); the run log says b10893 (c32d1dabe), which is wrong for these runs | -c 131072 -np 1 -fa on -ctk q8_0 -ctv q8_0 -ngl 999 --jinja --reasoning-format deepseek --alias qwen3.8-27b | extraction | the clean A/B against K2-36B: one variable (the weights), same flags, same slot count. The label records -np 1, because -np 2 to -np 1 had flipped 7 worldview and 8 nullschema cases at temperature 0 the day before |
qwen38-vendor-qwen-think | 2026-09-10 09:49–10:32 | b10618 (eb25b7263); run log says b10893 | same flags | vendor-qwen-think (thinking on, budget 8192, Qwen's recommended sampling) | layer 2 of the plan: each model at its own recommended settings, requested by Alan ("Qwen actually has recommended setting we should probably follow") |
qwen38-rebase-b10893 | 2026-09-10 12:24–12:29 | llama.cpp b10893 (c32d1dabe, 0.4.0-dev) | same flags | extraction | re-run of nullschema and worldview after Alan said "go upgrade": one commit in the 201 pulled touched the GBNF path nullschema depends on |
Seats held: the assistant's state summary at 12:29Z lists <port> qwen3.8-27b and its fleet table at 12:16Z gives the model 19.7 GB. The seat then moved to the Q4_0 build on 2026-09-11. Exact stop time of the server is not on record.
Decisions
- Pre-registered expectation: none on record for the numbers; the A/B script header states the single variable and the caveat that K2 needs the fork binary while Qwen runs on the standard build.
- Kept as the Qwen3.8 reference through 2026-09-11 because it was the baseline for the Q4_0 comparison (the Q4_0
nullschemafixed/broken counts compare againstqwen38-rebase-b10893). - Vendor pass killed. The assistant argued the pass would take 12 hours to confirm a negative on two more suites and block the mid-range cohort; Alan: "kill it".
- Not published separately.
What went wrong
- Partial run: the vendor-qwen-think nullschema file (above).
- Missing flags (standing check): all three labels have
--jinjaand--reasoning-format deepseek; the vendor profile turns thinking on through the request, not a server flag.
What happened to it
Downloaded 2026-09-10 06:55Z → A/B, vendor pass, rebase 2026-09-10 → requantised to Q4_0 2026-09-11 13:21Z (Qwen3.8-27B Q4_0 (requantised here from the UD-Q4_K_M)) → deleted 2026-09-12 08:04Z. No RESTORE entry (the 20260911 script has no Qwen3.8 line). Re-downloadable from unsloth/Qwen3.8-27B-GGUF.