Ornith-1.5-35B-A3B — Huihui × Sangreal ICE GGUF

GGUF quantizations of huihui-ai's abliterated Ornith-1.5-35B-A3B, made with a Sangreal 36,025-chunk importance matrix, ICE 1.6 per-tensor recipes and ICE 2.0 post-quantization expert repair. Every file keeps the MTP draft layer. The ICE files carry peculiar-ragdoll's v22.5.0 Qwen-Sharp chat template with gbuzhf's arrow4.0 harness operator wrapper.


Architecture	Qwen-3.5-35B-A3B (qwen35moe): 40 layers, 256 routed experts (8 active) + shared expert, hybrid Gated DeltaNet / full attention, ~3B active
Extras	MTPv2 head (1 draft layer, `nextn_predict_layers = 1`); vision projector ships separately — use Ornith's own mmproj directly
Source	huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated
Chat template	Arrow 4.0 (Qwen-Sharp v22.5.0 base), embedded in every ICE file

⚠️ Uncensored. This is an abliterated checkpoint — refusal behaviour has been suppressed in the weights. Sandbox it at the OS level and control its network and code-execution access; with no refusal backstop, a prompt injection from a hostile page or third-party code has nothing to stop it.

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	0.0274	0.0137	0.0376	93.4 % / 97.1 % / 96.8 %	97.0
23G-ICEv2	22.52 GB	0.0313	0.0173	0.0409	92.9 % / 96.8 % / 96.5 %	96.6
19G-ICEv2	18.47 GB	0.0511	0.0298	0.0548	91.1 % / 95.8 % / 95.7 %	95.2
15G-ICEv2	14.51 GB	0.1162	0.0734	0.1001	86.6 % / 93.6 % / 93.6 %	91.2
12G-ICEv2	11.90 GB	0.2027	0.1277	0.1541	82.3 % / 91.4 % / 91.6 %	86.8

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

Demo — Coral Dragon Reef Rescue

A complete browser game — one self-contained index.html, no assets, no network calls — generated in a single shot by the 25G ICE 2.0 build (Ornith-1.5-35B-A3B-Huihui-Sangreal-MTP-25G-ICEv2.gguf) in about 1 h 30 min, running locally in the Nous Hermes desktop harness on Windows 11 with an RTX 3070 (8 GB VRAM), 32 GB RAM and no MTP.

Recommended Sampling Parameters

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

temperature: 0.6
top_p: 0.95
top_k: 20

Use

llama-server -m Ornith-1.5-35B-A3B-Huihui-Sangreal-MTP-19G-ICEv2.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).

Files

File	Size	Routed experts	What
`Ornith-1.5-35B-A3B-Huihui-Sangreal-MTP-12G-ICEv2.gguf`	12.19 GB	IQ2_XS	ICE 2.0 (11.90 GB without the MTP head)
`Ornith-1.5-35B-A3B-Huihui-Sangreal-MTP-15G-ICEv2.gguf`	14.81 GB	IQ3_XXS + IQ2_S	ICE 2.0 (14.51 GB without the MTP head)
`Ornith-1.5-35B-A3B-Huihui-Sangreal-MTP-19G-ICEv2.gguf`	18.76 GB	IQ4_XS + IQ3_S	ICE 2.0 (18.47 GB without the MTP head)
`Ornith-1.5-35B-A3B-Huihui-Sangreal-MTP-23G-ICEv2.gguf`	22.81 GB	Q4_K + Q5_K	ICE 2.0 (22.52 GB without the MTP head)
`Ornith-1.5-35B-A3B-Huihui-Sangreal-MTP-25G-ICEv2.gguf`	24.82 GB	Q5_K + Q4_K	ICE 2.0 (24.53 GB without the MTP head)

Credits

Ornith-1.5 by ornith-ai; abliterated checkpoint by huihui-ai (MIT). Sangreal corpus, ICE recipes and quantization by gbuzhf. Chat template by peculiar-ragdoll (Qwen-Sharp) with gbuzhf's arrow4.0.