You can deploy this model and start earning money today!

georgeis55/
Ornith-1.0-35B-MLX-oQ8-mtp

Name: Ornith-1.0-35B-MLX-oQ8-mtp
Availability: InStock

Views

No views yet

Ornith-1.0-35B-MLX-oQ8-mtp

This model was quantized using oQ (oMLX v0.4.5.dev1) mixed-precision quantization.

oQ8 MLX build of deepreinforce-ai/Ornith-1.0-35B with a quantization-matched MTP head for native speculative decoding on Apple Silicon.

Ornith-1.0-35B ships without MTP tensors. This build grafts a qwen3_5_moe MTP block (1 nextn layer, 785 tensors, ~488 MB) sourced from Qwopus3.6-35B-A3B

Architecture

Base: Qwen3.5 MoE
MTP: 1 nextn layer at mtp.* (fc, pre_fc_norm_hidden/embedding, norm, full MoE attention+experts), declared via num_nextn_predict_layers: 1
Quantization: oQ mixed-precision, 8-bit, group size 64, MLX safetensors

Limitations

MTP head is a verbatim graft, never trained on Ornith's hidden states. ~60-70% acceptance is the practical ceiling for verbatim grafts on this architecture.

georgeis55/Ornith-1.0-35B-MLX-oQ8-mtp

Ornith-1.0-35B-MLX-oQ8-mtp

Architecture

Limitations

georgeis55/
Ornith-1.0-35B-MLX-oQ8-mtp