Modelpublic

aimeri/spoomplesmaxx-jackdaw-35B-A3-i1-GGUF

sourceHugging Faceupdated 2mo agoView on Hugging Face

0likes90downloads

Model Card

spoomplesmaxx-jackdaw-35B-A3-i1-GGUF

Weighted / imatrix quants of `aimeri/spoomplesmaxx-jackdaw-35B-A3`.

Static quants are not provided: at every size that matters the imatrix quant of the same name is at least as good, so this repo is the whole set.

Provenance


llama.cpp	`b9222` (pinned; Qwen3.5 MoE conversion needs >= b9180)
Source	bf16 GGUF converted from the SFT checkpoint with `convert_hf_to_gguf.py`
imatrix	`llama-imatrix --parse-special`, 192 MB
Calibration	48 MB, 12.0M in-domain tokens (952 documents) rendered through this model's own chat template and truncated to exact 8192-token windows so imatrix chunks land on document boundaries; the generic term is unsloth's published imatrix for this architecture, merged after our own pass

The in-domain half exists because this is a sparse MoE: a token only touches its routed experts, so a small generic corpus leaves rare experts with no importance data at all. Coverage was gated before quantizing — every quantizable tensor has importance data and every expert row was activated above a token floor.

The Q4_K_M quant was verified after quantization: greedy generations terminate on <|im_end|> (not the token cap), think blocks close, no stray glyphs.

Measured

KL divergence of each quant against this model's own bf16, on a holdout that was excluded from calibration (evaluating on your own calibration domain flatters the numbers). Lower KLD is closer to the unquantized model.

quant	size	KLD mean	KLD median	KLD p99	ΔPPL	same top-1
`i1-IQ4_XS`	18.7 GB	0.0223	0.0123	0.129	+0.83%	100.0%
`i1-Q4_K_M`	21.4 GB	0.0195	0.0102	0.121	+0.92%	100.0%

same top-1 is the share of holdout positions where the quant's most likely token is unchanged from bf16.

Which one do I want?

Roughly, higher in the table is smaller and worse. i1-Q4_K_M is the recommended default; i1-IQ4_XS if you need to save a little; i1-Q6_K if you have the memory and want to stop thinking about it.

Link	Type	Size (GB)	Notes
GGUF	i1-IQ4_XS	18.7
GGUF	i1-Q4KM	21.4	fast, recommended

Usage

bash

llama-server -m spoomplesmaxx-jackdaw-35B-A3-i1-GGUF.i1-Q4_K_M.gguf -ngl 99 -c 32768 --jinja

--jinja matters: the chat template is embedded in the GGUF and pre-opens the <think> block, which is the behaviour the model was trained for. Tool calls use Qwen3.5's XML convention (<tool_call><function=NAME><parameter=KEY>VALUE</parameter></function></tool_call>).