dongnhdev/Ornith-1.5-122B-MTP-GGUF

Ornith-1.5-122B — IQ2 + MTP (GGUF)

GGUF quantization of the Ornith-1.5-122B checkpoint (general.name = Qwen3.5-122B-A10B-ck2100-v1014-MTP1740, architecture qwen35moe, 899 tensors, 256 experts × 6.7B). The MTP head is included as blk.48.nextn.*, so multi-token-prediction draft modules work out of the box.

Files

file	size	notes
Ornith-1.5-122B-IQ2-MTP.gguf	37.36 GB	mixed IQ2_XXS / Q4_K / Q6_K / Q2_K, ~2.45 bpw, imatrix-assisted, MTP head included
mmproj-Ornith-1.5-122B-BF16.gguf	912 MB	vision projector (BF16)
mmproj-Ornith-1.5-122B-F16.gguf	909 MB	vision projector (F16)
imatrix_unsloth.gguf_file	359 MB	importance matrix used for the quant
mask_row_248077.f32	12 KB	vision-mask test row (reference fixture)

Quantization profile

Tensor type census of the main GGUF: F32 ×370 (norms/biases), Q4_K ×331, IQ2_XXS ×98 (routed experts), Q6_K ×51, Q2_K ×49 — weighted ~2.45 bits per weight. Routed experts use IQ2_XXS with an imatrix; attention/FFN dense tiles stay at Q4_K/Q6_K.

Usage

# text
llama-cli -m Ornith-1.5-122B-IQ2-MTP.gguf -p "Hello"

# vision (projector must match the runtime)
llama-mtmd-cli -m Ornith-1.5-122B-IQ2-MTP.gguf \
  --mmproj mmproj-Ornith-1.5-122B-F16.gguf --image photo.jpg -p "Describe"

MTP: runtimes that support Qwen3.5-MoE MTP can enable the blk.48.nextn draft block for speculative decoding; otherwise the model still runs as a plain MoE decoder.

Notes

• Quantized locally (general.quantization_version = 2).
• Sizes above are exact byte counts verified against the source files.