Kolibri-1 78B-A3.5B · file← Kolibri-1 78B-A3.5B

MLX 4-bit

A community conversion for Apple's MLX, used for our first tests. Deleted the next day to make room for the Q4_0. Velaia conversion; study and board runs.

Size: 44.3 GB
Status: Deleted 5 Oct 2026
Scored runs: 0

ResultsNever scored

This file never completed a scored run. Its record below says why.

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	9 safetensors shards (41.3 GiB)
Format	MLX 4-bit g64
Where from	downloaded 2026-10-04
Status	shards deleted 2026-10-05; shim/config kept
Hugging Face	`velaia/Kolibri-1-MLX-4bit`

Where it came from

Source: velaia/Kolibri-1-MLX-4bit (a community conversion; a nativ-community copy was also looked at). Nine safetensors shards, 41.3 GiB, plus run.py/kolibri1.py shim, kolibri1-llama.cpp.patch, chat_template.jinja (models/kolibri-1-mlx-4bit/). Revision and sha256 not on record; the manifest fingerprint is dir_fingerprint over all shards.
Quantisation: 4-bit, group size 64, affine; embeddings and lm_head 8-bit; router unquantised.
Download: download.sh with pardl.sh, 6 ranges per shard, nine shards sequentially, started 2026-10-04 18:41Z, ALL DONE 19:03Z (download.log: 20:03:26 local).
Why this file: the question prompt of 2026-10-04 offered MLX 4-bit (recommended) vs llama.cpp Q4_K_M; the chosen option is listed as "MLX 4-bit" in a later summary. Reasons: ~1.45x faster on this box per the server runbook, release mlx-lm, and the uploader's own numbers came from an M1 Max 64 GB.

How it was served

No run-log label exists for this variant: it was run on the v1 eval harness (eval/runner.py, results/kolibri-1-78b__study-default.*) and the 16-suite board (Downloads/benchmarks), neither of which writes the locked-harness run log.

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / --jinja / --reasoning-format	harness profile	why this label exists
`kolibri-1-78b__study-default` (v1 results, not in the run log)	smoke 2026-10-04 19:05Z; full 624-task run launched 19:17Z, finished 2026-10-05 04:39Z (05:39 local; "7 h 03 min" wall per the lab's note)	`mlx_lm` 0.32.0 on mlx 0.32.3 in `mlx-venv`, via the `run.py` shim	`--decode-concurrency` 4 on the smoke, 1 afterwards; no `-fa`/`--jinja` (not llama.cpp)	`study-default` (thinking on, 10k cap, temp 0), `configuration_id` `8a98d90205c54f14`	the 624-task study: first complete thinking-on run on that harness
grounding suite (v1)	2026-10-05 04:41-04:55Z	same	same, 4000-token cap	v1 `grounding.py`	48-case grounding
16-suite board `high`, then `low`	2026-10-05 06:30-08:24Z (high); ~08:33-09:34Z (low)	same server	`reasoning_effort` high / low	board harness

Seats held: none; the five launchd seats were unloaded for each window with eval/tools/seats.sh and a deadline guard.

Decisions

Chosen first over GGUF for speed-of-start and rule compliance (mlx-always-latest), not for final speed; the llama.cpp Q4_0 later beat it by 47%.
Shards deleted 2026-10-05 10:06Z to stage the Q8_0 -> Q4_0 work (approved by Alan). The shim, patch and config were kept so the MLX copy can be re-pulled in ~25 minutes if wanted.
Not published in the locked board; the 624-task numbers live in the lab's notes only.

What went wrong

mlx_lm batching was mis-generalised. The 4-way result (38.7 aggregate vs 48 single) was recorded as "fine-grained MoEs do not batch on Metal" and corrected on 2026-10-09 as an mlx_lm property (a lab note changelog).
Grader bug found by this run: graders._paras split IFEval paragraphs on blank lines instead of ***; fixed 2026-10-04, which also moved qwen3.6's and gemma4's IFEval numbers.
Run interruptions (all operator-side): output directory missing (smoke attempt 1), squish-venv had lost pyarrow (attempt 2), a deliberate queue reorder to cheap-first, and the 2-hour background-command cap killing the runner once; every restart resumed on the same configuration_id.
Missing flags (standing check): thinking was on at the template-default high effort for the study, which is the vendor's leaderboard setting and not a serving setting.
Stamp check: the manifest stamps mlx/mlx-lm versions, not a llama.cpp build; no llama_build is attached to these results (the check for MLX rows carrying a llama.cpp stamp is moot here).

What happened to it

Downloaded 2026-10-04 18:41-19:03Z -> smoke, 624-task study, grounding, 16-suite board (two efforts) -> shards deleted 2026-10-05 10:06Z; shim/config remain (du 9.3 MB). No RESTORE script entry. Re-downloadable from velaia/Kolibri-1-MLX-4bit via models/kolibri-1-mlx-4bit/download.sh.