NANI-Nithin/Ornith-1.5-35B-A3B-GGUF

Ornith-1.5-35B-A3B GGUF

GGUF quantizations of ornith-ai/Ornith-1.5-35B-A3B for use with llama.cpp, KoboldCpp, LM Studio, Open WebUI, Jan, and other GGUF-compatibledel

Ornith-1.5-35B-A3B is a Qwen3.5-MoE vision-language model featuring:

• 40 transformer layers
• Hybrid attention architecture combining:
  • Gated DeltaNet linear attention (3 out of every 4 layers)
  • Full attention layers
• Mixture-of-Experts architecture
  • 256 routed experts
  • 8 experts active per token
• 1 MTP layer
• MRoPE positional encoding
• Hugging Face architecture:
  • Qwen3_5MoeForConditionalGeneration
  • model_type: qwen3_5_moe

The upstream llama.cpp runtime supports the model family natively, including conversion, inference, importance matrix generation, and quantization.

Note: These GGUF files are text-only. The vision tower present in the original VLM checkpoint is skipped during GGUF conversion, which is standard practice for current llama.cpp VLM conversions.

Conversion Notes

This repository was generated directly from the original Hugging Face checkpoint using the current upstream llama.cpp conversion pipeline.

Because the original safetensors checkpoint is approximately 72 GB, generating a BF16 GGUF would require more temporary disk space than was available on the build machine.

Instead:

1. The checkpoint was converted directly to a Q8_0 GGUF
2. The Q8_0 file was uploaded as the highest-quality quant
3. All other quantizations were generated from that Q8_0 source using llama-quantize

This approach is standard in the llama.cpp ecosystem and produces effectively lossless source material for subsequent K-Quant and IQ-Quant generation.

Quantizations

Standard Quants

Quant	Recommended Use
Q2_K	Maximum compression
Q3_K_S	Very low-memory inference
Q3_K_M	Budget quality/runtime balance
Q3_K_L	Best Q3 variant
Q4_0	Legacy compatibility
Q4_1	Legacy quality-focused quant
Q4_K_S	Small Q4_K variant
Q4_K_M	Recommended general-purpose quant
Q5_K_S	Compact high-quality quant
Q5_K_M	Excellent quality/size tradeoff
Q6_K	Near-Q8 quality
Q8_0	Highest quality quant in this repo

IQ Quants

Generated using a llama.cpp importance matrix.

Quant
IQ2_M
IQ3_XXS
IQ3_XS
IQ3_M
IQ4_XS
IQ4_NL

Importance Matrix

IQ quantizations were generated using a llama.cpp importance matrix built from WikiText-2 calibration data.

To accommodate the build machine's memory constraints:

• Imatrix generation was performed on a temporary Q2_K model
• WikiText-2 was used as calibration data
• The resulting imatrix was then applied during IQ quantization

Recommended Files

Largest Quality

Q8_0

Use if storage and RAM are not concerns and you want the best quality available in this repository.

Best Overall

Q4_K_M

Recommended for most users.

Provides an excellent balance between quality, speed, memory usage, and disk footprint.

Higher Quality

Q5_K_M

Recommended for users with additional RAM who want quality closer to Q6/Q8 while keeping storage reasonable.

Maximum Compression

Q2_K

For extremely constrained systems where fitting the model matters more than quality.

llama.cpp Usage

Example:

llama-cli \
  -m Ornith-1.5-35B-A3B-Q4_K_M.gguf \
  -c 8192

Server:

llama-server \
  -m Ornith-1.5-35B-A3B-Q4_K_M.gguf

Compatibility

These files should work in:

• llama.cpp
• LM Studio
• KoboldCpp
• Open WebUI
• Jan
• Text Generation WebUI (llama.cpp backends)
• Any GGUF-compatible runtime with support for Qwen3.5-MoE architectures

For best compatibility, use a recent build of llama.cpp or a frontend based on a recent llama.cpp release.

Credits

Original Model

• ornith-ai for Ornith-1.5-35B-A3B

GGUF Conversion & Quantization

• Generated using upstream llama.cpp
• Quantized and uploaded by NANI-Nithin

Disclaimer

This repository contains quantized derivatives of the original model. All model behavior, training data, capabilities, and limitations originate from the base model authors.

Please review the original model card for licensing terms, intended use, limitations, and safety considerations.