Sage-A1.5 — Sangreal ICE GGUF

GGUF quantizations of InternScience/Sage-A1.5, made with a Sangreal 10k importance matrix, ICE 1.6 per-tensor recipes and ICE 2.0 post-quantization expert repair. Every file keeps the MTP draft layer; the vision projector ships separately (InternScience's F16, plus a Q8_0). The ICE files carry the Arrow 4.0-sage chat template.


Architecture	Qwen3.6-35B-A3B (qwen35moe): 40 layers, 256 routed experts (8 active) + shared expert, hybrid Gated DeltaNet / full attention, ~3B active
Extras	MTP head (1 draft layer, `nextn_predict_layers = 1`), vision projector (`mmproj`)
Source	InternScience/Sage-A1.5-F16-GGUF (master), InternScience/Sage-A1.5-Q8_0-GGUF
Chat template	Arrow 4.0-sage, embedded in every ICE file; Sage-A1.5's verification switch is kept (`--chat-template-kwargs '{"verification": true}'`, needs a system message)

Fidelity of the ICE 2.0 tiers (vs the F16 master)

Tier	Size (no MTP)	wiki 2K KLD	code 2K KLD	code 16K KLD	same top-1 (wiki / code / code 16K)	Score
25G-ICEv2	24.53 GB	0.0079	0.0067	0.0250	96.2 % / 98.1 % / 97.6 %	98.3
23G-ICEv2	22.52 GB	0.0105	0.0090	0.0275	95.7 % / 97.7 % / 97.4 %	98.0
19G-ICEv2	18.47 GB	0.0198	0.0166	0.0347	94.3 % / 96.9 % / 96.8 %	97.2
15G-ICEv2	14.51 GB	0.0510	0.0478	0.0738	90.8 % / 94.8 % / 94.9 %	94.3
12G-ICEv2	11.90 GB	0.1123	0.0922	0.1213	86.7 % / 92.6 % / 92.9 %	90.4

Score: F16 = 100; 70 % full-distribution fidelity 1/(1+KLD) + 30 % same-top-1, averaged over the three sets.

Recommended Sampling Parameters

For the best generation quality, we recommend the following sampling parameters:

temperature: 1.0
top_p: 0.95
top_k: 20
min_p: 0.0
presence_penalty: 1.5
repetition_penalty: 1.0

Use

llama-server -m Sage-A1.5-19G-ICEv2.gguf --mmproj mmproj-Sage-A1.5-F16.gguf --jinja -c 65536 -fa on

The MTP head is in every file; on a llama.cpp build with MTP support, enable it for speculative decoding (draft depth 1). A Q8_0 of the projector (mmproj-Sage-A1.5-Q8_0.gguf) can be used in place of the F16 one.

Files

File	Size	Routed experts	What
`Sage-A1.5-12G-ICEv2.gguf`	12.19 GB	IQ2_XS	ICE 2.0 (11.90 GB without the MTP head)
`Sage-A1.5-15G-ICEv2.gguf`	14.81 GB	IQ3_XXS + IQ2_S	ICE 2.0 (14.51 GB without the MTP head)
`Sage-A1.5-19G-ICEv2.gguf`	18.76 GB	IQ4_XS + IQ3_S	ICE 2.0 (18.47 GB without the MTP head)
`Sage-A1.5-23G-ICEv2.gguf`	22.81 GB	Q4_K + Q5_K	ICE 2.0 (22.52 GB without the MTP head)
`Sage-A1.5-25G-ICEv2.gguf`	24.82 GB	Q5_K + Q4_K	ICE 2.0 (24.53 GB without the MTP head)
`Sage-A1.5-Q8_0.gguf`	37.80 GB	Q8_0	InternScience's Q8_0, unchanged
`mmproj-Sage-A1.5-F16.gguf`	0.90 GB	—	Sage's vision projector, unchanged
`mmproj-Sage-A1.5-Q8_0.gguf`	0.61 GB	—	the same projector in Q8_0

Credits

Sage-A1.5 by InternScience (Apache-2.0). Sangreal corpus, ICE recipes and quantization by gbuzhf.

gbuzhf/Sage-A1.5-Sangreal-ICE