Apex-35B-V1.0 — Sangreal ICE GGUF

GGUF quantizations of Apexintelligence-AI/Apex-35B-V1.0, 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 (F16, plus a Q8_0). Every file carries the Arrow 4.0-apex chat template. An uncensored variant is at gbuzhf/Apex-35B-V1.0-Uncensored-Sangreal-ICE.


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	Apexintelligence-AI/Apex-35B-V1.0-GGUF (official BF16)
Chat template	Arrow 4.0-apex, embedded in every file (BF16, Q8_0, ICE tiers); Apex's own `<think>` opener is kept

Fidelity of the ICE 2.0 tiers (vs the BF16 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	0.0618	0.0284	0.0859	91.4 % / 96.2 % / 94.8 %	94.4
23G-ICEv2	22.52 GB	0.0796	0.0390	0.0976	90.4 % / 95.4 % / 94.2 %	93.3
19G-ICEv2	18.47 GB	0.1331	0.0684	0.1474	87.6 % / 94.0 % / 92.5 %	90.2
15G-ICEv2	14.51 GB	0.3067	0.1765	0.2965	81.7 % / 90.4 % / 89.2 %	81.8
12G-ICEv2	11.90 GB	0.5248	0.2957	0.4601	76.4 % / 87.9 % / 86.6 %	74.4

Score: BF16 = 100; 70 % full-distribution fidelity 1/(1+KLD) + 30 % same-top-1, averaged over the three sets.

[!Note] Apex-35B-V1.0 is unusually sensitive to rounding: even a plain Q8_0 of the original, un-abliterated model is ~0.015 KLD from its BF16 on code, so these values run higher than other 35B-A3B ladders at the same size. That is a property of the base model, not of the ICE quantization.

Recommended Sampling Parameters

From Apex-35B-V1.0's own card:

temperature: 1.0
top_p: 0.95
top_k: 20
max_tokens: 65536 (the output budget per reply, thinking included)

Use

llama-server -m Apex-35B-V1.0-19G-ICEv2.gguf --mmproj mmproj-Apex-35B-V1.0-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-Apex-35B-V1.0-Q8_0.gguf) can be used in place of the F16 one.

Files

File	Size	Routed experts	What
`Apex-35B-V1.0-12G-ICEv2.gguf`	12.19 GB	IQ2_XS	ICE 2.0 (11.90 GB without the MTP head)
`Apex-35B-V1.0-15G-ICEv2.gguf`	14.81 GB	IQ3_XXS + IQ2_S	ICE 2.0 (14.51 GB without the MTP head)
`Apex-35B-V1.0-19G-ICEv2.gguf`	18.76 GB	IQ4_XS + IQ3_S	ICE 2.0 (18.47 GB without the MTP head)
`Apex-35B-V1.0-23G-ICEv2.gguf`	22.81 GB	Q4_K + Q5_K	ICE 2.0 (22.52 GB without the MTP head)
`Apex-35B-V1.0-25G-ICEv2.gguf`	24.82 GB	Q5_K + Q4_K	ICE 2.0 (24.53 GB without the MTP head)
`Apex-35B-V1.0-Q8_0.gguf`	37.80 GB	Q8_0	pure Q8_0 of the BF16 master
`BF16/Apex-35B-V1.0-BF16-0000{1,2}-of-00002.gguf`	44.87 + 26.20 GB	BF16	Apex's official BF16, with the Arrow 4.0-apex template stamped in (tensors unchanged)
`mmproj-Apex-35B-V1.0-F16.gguf`	0.90 GB	—	Apex's vision projector (F16), unchanged
`mmproj-Apex-35B-V1.0-Q8_0.gguf`	0.61 GB	—	the same projector in Q8_0
`imatrix/Apex-35B-V1.0-sangreal-10000.imatrix.gguf`	0.19 GB	—	the Sangreal 10k importance matrix used for every tier

Credits

Apex-35B-V1.0 by Apex Intelligence (Apache-2.0). Vision projector conversion by bartowski. Sangreal corpus, ICE recipes and quantization by gbuzhf.

gbuzhf/Apex-35B-V1.0-Sangreal-ICE