mudler/Ornith-1.5-35B-A3B-APEX-MTP-GGUF

⚡ Each donation = another big MoE quantized

I host 30+ free APEX MoE quantizations as independent research. My only local hardware is an NVIDIA DGX Spark (122 GB unified memory), enough for ~30-50B-class MoEs, but bigger ones (200B+) require rented compute on H100/H200/Blackwell, typically $20-100 per quant.
If APEX quants are useful to you, your support directly funds those bigger runs.

🎉 Patreon (Monthly)  |  ☕ Buy Me a Coffee  |  ⭐ GitHub Sponsors

Ornith-1.5-35B-A3B APEX MTP GGUF

APEX quantizations of ornith-ai/Ornith-1.5-35B-A3B.

Brought to you by the LocalAI team | APEX Project

These files bundle the model's MTP / NextN draft head as blk.40, so speculative decoding runs against the file itself with --spec-type draft-mtp. For the same quants without the head, see Ornith-1.5-35B-A3B-APEX-GGUF.

Files

File	Size	For
Ornith-1.5-35B-A3B-APEX-MTP-Quality.gguf	23.72 GB	highest quality
Ornith-1.5-35B-A3B-APEX-MTP-Balanced.gguf	26.17 GB	general purpose
Ornith-1.5-35B-A3B-APEX-MTP-Compact.gguf	17.44 GB	consumer GPUs
Ornith-1.5-35B-A3B-APEX-MTP-I-Mini.gguf	14.37 GB	smallest, imatrix only
mmproj.gguf	0.90 GB	vision projector, pair with any of the above

I- files use an importance matrix built from diverse calibration data (chat, code, reasoning, tool-calling, agentic traces, Wikipedia). Quality, Balanced and Compact also ship without it.

The model

Ornith-1.5-35B-A3B is a 36 B parameter Mixture-of-Experts model with 256 routed experts and 8 active per token, plus a shared expert. It has 40 layers with hybrid attention, interleaving three linear-attention layers per full-attention layer, and a vision tower.

How APEX quantizes it

Routed experts are 89.6% of the weights here but only 8 of 256 fire for any given token, so they tolerate lower precision than the parts every token passes through. APEX classifies each tensor by role and applies a layer-wise precision gradient: the first and last layers keep higher precision, middle layers compress harder, and the always-active shared expert is kept high.

Attention is only 3.6% of the weights on this model (2.8% linear, 0.8% full), so it is not where the size is and is not treated as a lever.

The MTP head is a full MoE block in its own right, about 2.4% of the weights. On Quality, Balanced and Compact it is pinned to Q8_0, since a drafter that mispredicts the target wastes the speculation it was added for. I-Mini keeps it at tier precision to stay small.

Usage

# text
llama-cli -m Ornith-1.5-35B-A3B-APEX-MTP-Balanced.gguf -p "Your prompt" -ngl 99

# vision
llama-mtmd-cli -m Ornith-1.5-35B-A3B-APEX-MTP-Balanced.gguf --mmproj mmproj.gguf -ngl 99

# speculative decoding against the bundled MTP head
llama-cli -m Ornith-1.5-35B-A3B-APEX-MTP-Balanced.gguf --spec-type draft-mtp -ngl 99

Needs a recent llama.cpp with qwen3_5_moe support.

Notes

Sizes and quantization recipes are published in the APEX repository. No throughput benchmarks were run on these files.