spoomplesmaxx-whiskeyjack-12B-i1-GGUF

Weighted (imatrix) GGUF quants of aimeri/spoomplesmaxx-whiskeyjack-12B, quantized on the box that trained it.

Measurements

KL divergence against this model's own bf16, on held-out text that was excluded from the calibration corpus by construction. Not against the base model, not against a benchmark — the number answers one question: how much did quantization change this model.

file	GB	KLD mean	KLD median	KLD p99	RMS Δp %	same top-1 %
spoomplesmaxx-whiskeyjack-12B.i1-IQ2_M.gguf	4.80	0.2262	0.1187	1.8559	14.50	81.03
spoomplesmaxx-whiskeyjack-12B.i1-IQ3_M.gguf	6.17	0.0545	0.0252	0.4786	7.14	90.44
spoomplesmaxx-whiskeyjack-12B.i1-IQ4_XS.gguf	7.17	0.0287	0.0114	0.2847	5.52	93.46
spoomplesmaxx-whiskeyjack-12B.i1-Q4_K_M.gguf	7.95	0.0221	0.0088	0.1982	5.01	94.21
spoomplesmaxx-whiskeyjack-12B.i1-Q5_K_M.gguf	9.24	0.0102	0.0034	0.1006	3.63	96.42
spoomplesmaxx-whiskeyjack-12B.i1-Q6_K.gguf	10.61	0.0038	0.0012	0.0366	2.16	97.93

Calibration

context	4096 tokens, document-aligned
corpus	11.8M tokens, 2093 documents
sources	100% in-domain (the model's own training corpora)
separator	<eos> (a special token — see below)
reference imatrix	merged from unsloth/gemma-4-12b-it-GGUF

Every document is truncated to an exact multiple of the calibration context, so llama-imatrix's non-overlapping windows land on document boundaries rather than straddling two unrelated scenes. The separator is a special token because whitespace separators are BPE-mergeable: a document ending in a newline and the next beginning with one can fuse into a single token and shift every subsequent window.

CALIB_CTX is 4096 rather than the 8192 used for larger models in this family, and that is measured rather than inherited: at 8192 only 0.1% of the aviary corpus clears a single window, which would have made calibration ~95% one sub-corpus with no tool-calling coverage at all. Gemma 4 also runs 40 of its 48 layers as sliding-window attention at 1024, so context beyond a few thousand tokens sharpens statistics for only the 8 global layers.

Serving

The stop token is <turn|> (id 106), not <eos>. generation_config.json carries a list and GGUF stores a single u32; this build is verified to have picked <turn|>. Anything that waits for <eos> will run past the turn.

Thinking is selected by <|think|> at the top of the system turn. With it, every model turn opens a <|channel>thought ... <channel|> block; without it, there is no channel at all.

For tool use, a whole episode lives inside ONE <|turn>model: serve with stop=["<tool_call|>"], inject <|tool_response>response:NAME{...}<tool_response|>, and continue the same turn. A harness waiting for <turn|> after a tool call will hang.