Agents-A1 35B · file← Agents-A1 35B
Q4_K_M
The one we kept: InternScience's own Q4_K_M, and the file every test ran on. InternScience GGUF.
- Size
- 21.2 GB
- Status
- On disk
- Scored runs
- 7
- Smallest Mac
- 36 GB estimate, 8K context
Results7 scored · 95% intervals
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| epistemicsagents-a1-35b | 100.0% | 92.6–100.0 | Passed 48/48 | llama.cppstudy-nothink | 2026-08-24 |
| groundingagents-a1-35b | 71.2% | 62.7–78.4 | Passed 89/125 | llama.cppextraction | 2026-08-24 |
| groundingagents-a1-35b-b10618 | 71.2% | 62.7–78.4 | Passed 89/125 | llama.cppextraction | 2026-08-26 |
| nullschemaagents-a1-35b-b10618 | 38.9% | 28.5–50.4 | Passed 28/72 | llama.cppextraction | 2026-08-26 |
| screeningagents-a1-35b-b10618 | 95.0% | 86.3–98.3 | Passed 57/60 | llama.cppextraction | 2026-08-26 |
| toolsagents-a1-35b | 100.0% | 79.6–100.0 | Passed 15/15 | llama.cppagentic-fast | 2026-08-24 |
| toolscaleagents-a1-35b | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppagentic-fast | 2026-08-24 |
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.
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 | Agents-A1-Q4_K_M.gguf (21,166,757,632 B) |
| Format | GGUF Q4_K_M (general.file_type 15) |
| Where from | downloaded 2026-08-12 |
| Status | benched; router preset [agents-a1-35b]; on disk 2026-10-23 |
| Hugging Face | InternScience/Agents-A1-Q4_K_M-GGUF |
Where it came from
- Repo and file:
InternScience/Agents-A1-Q4_K_M-GGUF,Agents-A1-Q4_K_M.gguf, fetched byfleet-roster/models-scripts/get-agents-a1.shwithpget.sh(6 parallel range requests) fromhttps://huggingface.co/InternScience/Agents-A1-Q4_K_M-GGUF/resolve/main. - Date: file mtime 2026-08-12 17:21 local (BST). Completion marker
.dl_complete17:21. - Revision and sha256: not on record. The download did not use
hf download, so no HF metadata was left, and no checksum was written (would have been in the download log, not kept). - Who asked: not on record (the 2026-08-12 session is not available).
- GGUF header (read 2026-10-23): architecture
qwen35moe,file_type15 (Q4_K_M), 733 tensors, GGUF v3, 40 blocks, 256 experts / 8 used,context_length262144. - The companion projector is Agents-A1 35B vision projector (mmproj, F16).
How it was served
All launches used the plain upstream llama-server at llama.cpp/build/bin/llama-server, --jinja, -ngl 99, -cb, KV cache type and -fa left at defaults (f16 / auto), no --reasoning-format on the line, no --mmproj, no draft model. Run-log runtime_source is evidence for every entry.
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
agents-a1-35b (tools, epistemics) | 2026-08-24 00:04-00:05Z (run-log 01:04:20 and 01:05:04 BST); one-off script written 2026-08-23 23:44Z | llama.cpp b10369 (6e62ba538) | 32768 / 2 / f16 / auto / yes / not set | agentic-fast (thinking on, reasoning budget 512); study-nothink | the instrument check: does the suite discriminate? |
agents-a1-35b (toolscale) | 2026-08-24 01:17Z (run-log 02:17:44 BST); script 00:52Z | b10369 | 16384 / 2 / f16 / auto / yes / not set | agentic-fast | the staircase suite built because tools saturated |
agents-a1-35b (grounding) | 2026-08-24 13:07Z (run-log 14:07:08 BST); queued behind the Toolery sweep at 07:39Z | b10369 | 32768 / 4 / f16 / auto / yes / not set | extraction | grounding/1 sweep over the same six models as Toolery |
agents-a1-35b-b10618 (grounding, screening, nullschema) | script launched 2026-08-25 08:08Z, gated on a GPU lock; agents-a1 ran last, results 2026-08-26 07:32-07:35Z (run-log 08:32:13, 08:33:35, 08:34:56 BST) | llama.cpp b10618 (eb25b7263) | 32768 / 4 / f16 / auto / yes / not set | extraction | the fleet re-run on the rebuilt llama.cpp, tagged -b10618 so the two builds cannot be mixed |
Other servings of this file that are not run-log labels:
- Toolery (separate repo, not the locked harness):
--model agents-a1-35b,-c 32768 -np 4 -cb --jinja, 1 trial, 143 scenarios, started 2026-08-24 11:04Z. A first 3-trial attempt was killed at 412/429 traces. A-b10618Toolery run is in a one-off script; its result is not on record. - Router preset
[agents-a1-35b]infleet-models.ini:model =.../agents-a1-35b/Agents-A1-Q4_K_M.gguf,c = 131072, served by the router (--models-max 1). Present from before 2026-08-23 and still present on 2026-10-11. - Hermes seat, 2026-09-13 22:10Z to 2026-09-14 14:30Z:
llama-server --model models/agents-a1-35b/Agents-A1-Q4_K_M.gguf --jinja -c 65536 -np 1 --cache-ram 2048 -ctk q4_0 -ctv q4_0 -fa on -ngl 999 --alias agents-a1-35b. 49.1 tok/s, 21.4 GB resident. The committed copyfleet-roster/models-scripts/hermes-ab-35b.shis a later revision (-c 262144 -np 2 --cache-ram 16384), so do not take its flags as what ran.
Decisions
- Downloaded for the tuning-only A/B against Qwen3.6-35B-A3B (see the family page).
- Kept through every sweep: 2026-09-09 as "100% on three saturated suites", 2026-09-12 as an active seat that must not be deleted because no Q4_0 exists, and in the 2026-10-11 sweeps because the router preset still points at it.
- No Q4_0 requant, per
hermes-ab-35b.sh.
What went wrong
- Saturation, 2026-08-24. Tied a 3.35 GB model on tools and a 12B on epistemics; cost: the first routing table built on those suites. Fix:
toolscalestaircase (it also saturated at 64 tools), then Toolery. - Partial Toolery run, 2026-08-24. Killed at 412/429 and removed. An earlier mislabelling of the complete Gemma 12B run as abandoned was corrected in the same message; no agents-a1 data was published from the partial.
- Empty long-context results.
results_longctx/agents-a1-35b.jsonlhas 12 rows, all empty (reasoning consumed the token budget); reported as UNMEASURED, not as 0%. - Memory paging at the Hermes seat, 2026-09-13. With both 860 MB projectors loaded the pair ran at 11.5-14.9 tok/s with oscillating prefill; removed the projectors.
- Stray router instance, 2026-09-28. A client asked the preset router for this model at 00:11 local; with
--models-max 1the child stayed resident with a 131k unquantised KV cache and no house-ctk/-ctv q4_0flags. The assistant could not stop it and asked Alan to runkill 50671. On 2026-09-29 the router log tail shows "instance name=agents-a1-35b exited with status 1"; the cause is not on record. - Standing checks. Partial runs: none; all seven run-log entries have result-file rows equal to their attempted count (15, 48, 40, 125, 125, 60, 72). Stale build stamps: none found; the b10369 stamp on the 2026-08-24 runs matches the rebuild time. Missing flags:
--jinjawas on every line;--reasoning-formatwas never set.
What happened to it
Downloaded 2026-08-12 → benched on the v1 board 2026-08-21..23 → locked-harness runs 2026-08-24 and 2026-08-26 → router preset (c = 131072) → Hermes seat 2026-09-13 to 2026-09-14 → idle preset. Still on disk (models/agents-a1-35b/, 2026-10-23). Not deleted, so no RESTORE-deleted-*.sh entry. Re-downloadable from InternScience/Agents-A1-Q4_K_M-GGUF.