North-Mini-Code-1.0-43B-a5b
North-Mini-Code-1.0-43B-a5b is a 168-expert, top-14 sparse MoE expansion of
CohereLabs/North-Mini-Code-1.0 for coding and agentic software-engineering
workflows.
This upload is staged in an unfused Hugging Face-style MXFP4 layout. Expert
weights remain as per-expert tensors with MXFP4 block and scale tensors instead
of vLLM fused w13/w2 tensors.
Model Details
- Architecture: Cohere2 MoE causal language model
- Base model:
CohereLabs/North-Mini-Code-1.0 - Experts: 168 total, 14 active per token
- Quantization: MXFP4 MoE expert weights
- Layout: unfused per-expert MXFP4 safetensors index
- Intended use: code generation, terminal workflows, and tool-use experiments
Runtime Note
This layout is closer to the original Hugging Face checkpoint semantics than a
vLLM-fused export. Generic runtimes must understand the per-expert MXFP4
weight_blocks and weight_scales tensors. For native vLLM serving, use a
validated runtime conversion or a vLLM build that supports this unfused MXFP4
layout.
License
Released under the Apache 2.0 license, following the base model.