Per-group hybrid quantizations of ornith-ai/Ornith-1.5-35B-A3B (qwen3_5_moe, ~35B total / A3B active, vision-capable), produced by MagicQuant's measured evolutionary search: tensors are grouped by role, candidate per-group scheme assignments are rendered and perplexity-measured rather than predicted, and a winner is selected per size band. imatrix- calibrated (510 tensors), KL-blended (weight 0.1), 15 measured candidates over 3 rounds.

Pins: MagicQuant ed17991, Foundry 2f99202.

Files

Measured on one instrument: CPU llama.cpp perplexity, ctx 512, 100 chunks, wikitext-2 wiki.test.raw. BF16 baseline: 66.19 GiB, PPL 7.9691.

file	size	ratio vs BF16	PPL	vs baseline
`Ornith-1.5-35B-A3B-Q4_K_M.gguf`	19.58 GiB	0.30×	7.9530	−0.20%
`Ornith-1.5-35B-A3B-Q5_K_M.gguf`	22.88 GiB	0.35×	7.9951	+0.33%
`Ornith-1.5-35B-A3B-Q6_K.gguf`	28.84 GiB	0.44×	7.9792	+0.13%
`mmproj-Ornith-1.5-35B-A3B-f16.gguf`	0.86 GiB	—	—	vision projector

Each tier mixes schemes per tensor group rather than applying one quant type uniformly — the filenames name the size band, not a uniform type.

How to read these numbers honestly

Q4 measuring below the BF16 baseline does not mean it is better than BF16. At this instrument's resolution (ctx 512 / 100 chunks) a ±0.3% spread is noise, and the Q4 < Q6 < Q5 ordering among the three is not physically meaningful. The defensible claim is narrower and more useful:

Q4, Q5 and Q6 are indistinguishable from BF16 and from each other at this resolution. Separating them would need a sharper instrument (paired KL divergence, roughly 100× the resolution), which was not run.

The real signal is the cliff below these tiers. The same search also measured winners it did not render: Q3 (15.58 GiB) costs +2.96%, and Q2 (10.33 GiB) collapses at +30.2%. That is where quality actually starts being spent.

So: take Q4_K_M. It is 0.30× the BF16 size at a cost this instrument cannot separate from zero. The larger tiers are published for anyone who wants headroom, not because they are measurably better here.

Notes

Vision requires the mmproj file alongside the model.
Chat template embedded and verified present in every file before upload.
Vocab metadata (tokenizer.ggml.token_type) is INT32 per spec — these files load on current mainline llama.cpp builds.
MoE with a shared-expert split; expect the usual llama.cpp requirements for qwen3_5_moe. A build from mid-2026 or newer is recommended.
Built and measured on an AMD Strix Halo (Ryzen AI MAX+ 395, gfx1151) host.

lmcoleman/Ornith-1.5-35B-A3B-MagicQuant-GGUF

Ornith-1.5-35B-A3B — MagicQuant hybrid GGUFs

Files

How to read these numbers honestly

Notes

lmcoleman/
Ornith-1.5-35B-A3B-MagicQuant-GGUF