Devstral Small 2 24B · file← Devstral Small 2 24B
UD-Q4_K_XL
The file every score came from. A 4-bit copy, chosen so our tool tests would measure the model rather than its speed. Deleted in a tidy-up. Unsloth.
- Size
- 14.5 GB
- Status
- Deleted 7 Sep 2026
- Scored runs
- 3
- Smallest Mac
- 32 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| groundingdevstral-small-2-24b-b10618 | 79.2% | 71.3–85.4 | Passed 99/125 | llama.cppextraction | 2026-08-25 |
| nullschemadevstral-small-2-24b-b10618 | 62.5% | 51.0–72.8 | Passed 45/72 | llama.cppextraction | 2026-08-25 |
| screeningdevstral-small-2-24b-b10618 | 100.0% | 94.0–100.0 | Passed 60/60 | llama.cppextraction | 2026-08-25 |
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 | Devstral-Small-2-24B-Instruct-2512-UD-Q4_K_XL.gguf |
| Format | UD-Q4_K_XL GGUF, 14,506,151,072 B |
| Where from | downloaded 2026-08-24, sha256 verified against the LFS oid |
| Status | deleted 2026-09-07 17:36Z |
| Hugging Face | https://huggingface.co/unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF |
Where it came from
- Repo / file:
unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF/Devstral-Small-2-24B-Instruct-2512-UD-Q4_K_XL.gguf. - sha256:
b44e34b78180fc3ab1abbe1edad9f1f3926fdca10eed3bfae168b065e683f6cd, checked after assembly by a one-off script. The script's "OK" line for this file is not in the evidence excerpt; that it passed is implied by the script writing.dl_completeonly after both files verified, and by the Q6_K (second in the script) being present at 17:43 BST. - Revision: not on record (
resolve/main). - Date: 2026-08-24, between 16:15Z (download restarted) and 16:43Z.
- Method: six parallel range requests with
--speed-limit 51200 --speed-time 30, after a single-streamhf_hub_downloadcrawled at ~22 Mbps. - Who asked: Alan, "let's try out Devstral Small 2, what sizes does it come in?". The quant choice was the assistant's: "UD-Q4_K_XL at 14.5GB should decode in the 20–30 t/s range, which is enough to complete scenarios".
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
devstral-small-2-24b-b10618 | 2026-08-25 08:09Z (server up) – 08:37Z (last run-log row) | llama.cpp/build/bin/llama-server, b10618 (eb25b7263) | -c 32768 -np 4 -cb -ngl 99; KV default (f16); -fa not set; --jinja yes; --reasoning-format not set | extraction (thinking off, max_tokens 2048) | first measurable build; the -b10618 suffix keeps it from mixing with b10369 numbers |
| (not in run log) Toolery 143, same server | 2026-08-25 08:09Z–~08:30Z | b10618 | as above | Toolery --adapter raw --tier all --trials 1 --concurrency 4 | same fleet re-run, Toolery stage |
| (not in run log) AgentDojo banking + slack | 2026-08-25 20:43Z → finished by 2026-08-26 08:47Z | b10618 | -c 32768 -np 4 -cb -ngl 99 --jinja | AgentDojo --model VLLM_PARSED, utility then --attack important_instructions, logdir runs_full/devstral-small-2-24b | the injection-resistance question |
| (no label; gate abort) | 2026-08-24 ~19:2xZ | b10369 (6e62ba538) | -c 32768 -np 4 -cb -ngl 99 --jinja | — | a one-off script; aborted at the tool-call gate |
| (no label; gate abort) | 2026-08-24 19:31Z | b10369 | same | — | a one-off script with a GPU lock; aborted at the gate |
Run-log runtime for the scored label: b10618 (eb25b7263), runtime_source "evidence". The run-log llama_build stamp agrees. The endpoint is confirmed by all three result manifests. Never a seat, a router preset or a launchd plist.
Decisions
- Downloaded as the fleet-comparable 4-bit.
- Measured first in the b10618 re-run, because it and ornith were the reason for the rebuild.
- Not seated. "Right now it doesn't have a confirmed role". After AgentDojo, "12b is still the pick, and Devstral's role remains unproven". Named screener "if screening proves necessary" (a lab note §7).
- Deleted 2026-09-07 17:36Z as the smaller of two "duplicate quants", keeping the Q6_K, on Alan's "Go for it". The record does not show that the deletion weighed which file carried the scores.
- Published on the site as
devstral-small-2-24b-b10618(status "candidate").
What went wrong
- Tool-call gate aborts on b10369 (2026-08-24). Both gates failed with
JSONDecodeErroron"arguments":"{"location": "Warsaw"}". They were diagnosed as a llama-server serialisation bug, "model-specific — Gemma and Qwen serialise fine through the identical path". That diagnosis is the basis of the line "llama.cpp build 10369 emitted tool-call arguments with unescaped quotes, making Devstral and Ornith unmeasurable" (a lab note §5). What the record does not settle (reconstructed): - both b10369 gates piped the response through zsh
print; - the next morning, the same
print-based probe failed on b10618 for devstral, ornith and gemma4-e4b; - the cause turned out to be
printitself, which "interprets backslash escapes, so it mangled\""; - the file-based probe then passed on b10618;
- b10369 was never re-probed without
print, so whether it was really broken for Devstral is not on record.
Cost: one evening, and a full fleet re-run on a new build. Fix: gates write to a file.
- Scored file deleted, unscored sibling kept (2026-09-07). Afterwards, inventories reported devstral as an untested Q6_K: "no result file, no seat UNTESTED" and "NEVER BENCHMARKED". Fixed in
model_inventory.pyby token-set matching.
What happened to it
Downloaded 2026-08-24 → gate-blocked on b10369 2026-08-24 → benched on b10618 2026-08-25 (run log, Toolery, AgentDojo) → never seated → deleted 2026-09-07 17:36Z (duplicate-quant pass). No RESTORE entry (no RESTORE-deleted-20260907.sh exists; grep -i devstral over the RESTORE scripts finds only the 2026-10-11 Q6_K entry). Re-downloadable from unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF; verify against sha256 b44e34b7…83f6cd.