Coding-ablation-guided mixed-precision quantization of CohereLabs/North-Mini-Code-1.0 (cohere2moe, 128-expert MoE).

Cerebellum measures, per tensor group, what actually breaks the model's coding ability when crushed to Q2_K — using real HumanEval pass@1 deltas, not perplexity (perplexity is blind to code generation). It then writes one GGUF that protects the groups that matter and crushes the rest. On North, the per-group HumanEval ablation found exactly one coding-critical group — the routed down-projection experts — so v1 keeps those at Q4_K_M and crushes everything else to Q2_K.

Measured directly on this GGUF with llama.cpp (current master). --parallel 1, temperature 0.

North-Mini-Code is a reasoning-native model — it emits its full chain-of-thought in the response content (no <think> delimiter for llama.cpp to split on). A naive EvalPlus/HumanEval harness extracts the reasoning prose instead of the final code and massively under-scores the model (~51% on a stock extractor). The numbers above use a reasoning-aware extractor that recovers the model's actual final code block before execution; on inspection, the remaining failures are genuine logic errors plus a small number of reasoning-overflow cases, not extraction noise. If you bench this model, strip the reasoning to the final code block, or disable thinking (below).

Per the project's standing rule, comparisons are to the base model and same-size baselines only.

Built from the complete per-group HumanEval coding ablation (baseline = uniform Q4_K_M, 8 groups, real pass@1 deltas):

Norms protected (default). Base type Q4_K_M; only the groups above are overridden.

This is a cohere2moe model — it needs a llama.cpp build with cohere2-MoE support (merged to master, PR #24260 / commit 4988f6e). The GGUF uses the tiny_aya pre-tokenizer; older/PR-head builds that expect cohere2moe will fail to load. Build from current ggml-org/llama.cpp master.

Fits a 16 GB card. The model reasons in the content channel; for agent loops, pass the reasoning content forward between turns (per the base model's guidance).

Per-benchmark detail and the coding-ablation results are in benchmark_results/.