North-Mini-Code-1.0 GGUFs converted with ik_llama.cpp

These GGUF files are reviewer-facing and community-usable artifacts for CohereLabs North-Mini-Code-1.0 support in ik_llama.cpp.

They were converted from the official CohereLabs North-Mini-Code-1.0 checkpoint with the ik_llama.cpp Cohere2-MoE support added in PR #1945.

No performance or quality claims are made for these files. Local validation only confirms that the listed GGUFs load and respond on the tested hardware/runtime.

Files

Source And Build Notes

• Source model: CohereLabs/North-Mini-Code-1.0
• Source revision used locally: effaeda477c041c107d5a3d8c599cb5d6c5878ef
• Converter/runtime support: ik_llama.cpp main after PR #1945, or an equivalent build with cohere2_moe support
• Build commit used for these artifacts: 1e063a6bd Enhance Cohere2-MoE support by modifying tensor handling and configuration logic
• Conversion intermediate: BF16 GGUF, then quantized with llama-quantize
• Embedded chat template: source chat_template.jinja
• Embedded chat template SHA256 prefix: d8366efb9f07c571
• Tokenizer pre-tokenizer marker: tokenizer.ggml.pre=cohere2_moe

Relevant metadata present in all listed GGUFs:

• general.architecture=cohere2_moe
• tokenizer.ggml.pre=cohere2_moe
• tokenizer.chat_template is embedded and matches the source chat_template.jinja

Validation

Validation was run locally on an RTX 4070 workstation with an ik_llama.cpp build containing Cohere2-MoE / North-Mini-Code support.

Checks performed:

• BF16 conversion completed successfully.
• Q8_0, Q6_K, and Q4_K_M quantization completed successfully.
• Metadata was inspected after quantization.
• The embedded chat template in Q8_0, Q6_K, and Q4_K_M exactly matched the source chat_template.jinja.
• Q8_0, Q6_K, and Q4_K_M loaded through llama-server with the Cohere2-MoE runtime path.
• Q8_0, Q6_K, and Q4_K_M passed a short OpenAI-compatible chat sanity check.
• Q8_0, Q6_K, and Q4_K_M were also exercised through a local OpenCode configuration against an OpenAI-compatible endpoint and produced code in that harness.

These checks are compatibility smokes only. They are not benchmark results and should not be interpreted as performance, quality, or agentic coding claims.

Compatibility

These files require ik_llama.cpp main after PR #1945, or another runtime with equivalent cohere2_moe architecture, tensor-loading, tokenizer, and graph support.

They are not expected to load or decode correctly in runtimes that do not yet understand general.architecture=cohere2_moe.