Nanbeige4.2-3B Heretic GGUF

Selected 4-bit-and-up GGUF quantizations of WaveCut/Nanbeige4.2-3B-heretic.

File	Role
`Nanbeige4.2-3B-heretic-Q4_K_M.gguf`	Recommended compact mixed-precision default
`Nanbeige4.2-3B-heretic-Q5_K_M.gguf`	Recommended quality/speed balance
`Nanbeige4.2-3B-heretic-Q6_K.gguf`	High-quality K-quant
`Nanbeige4.2-3B-heretic-Q8_0.gguf`	Near-lossless quality anchor

No legacy Q4_0, Q5_0, or redundant same-bit variants are included.

The K-quants were calibrated with an importance matrix built from a deterministically shuffled agentic/coding corpus. Sources and revisions:

TIGER-Lab/SWE-QA-Pro-SFT-Trajectories at b8f5b8a8dcf90bca8b6d70adedac0d20dca02b86.
nvidia/OpenCodeReasoning at 20a1ca19c0d050fe9057fc08339d6b370ec1c67a.

Corpus SHA-256: a7cfdbe02c124304bf1282bbd5ed7162bfa72dec6750b60ed2d3a68000c7a554. The imatrix input takes 256 evenly spaced corpus records, truncates each to 1,024 tokens, and processes 256 context-1,024 input chunks with special-token parsing. Nanbeige's two execution loops produce 512 internal imatrix passes. The derived text SHA-256 is de99993785d460f0c48a8c35d36b764ef217ee227d6354cbeae0dc2cb155a30e.

Compatibility

Nanbeige 4.2 is a looped Transformer: 22 physical layers are executed twice. These files were converted and validated with Nanbeige's llama.cpp branch at revision 26cfdc4409cfc67d27be9b71c9de79adaf5f306f. Use that revision or a newer llama.cpp build containing equivalent Nanbeige support.

./llama-cli \
  -m Nanbeige4.2-3B-heretic-Q5_K_M.gguf \
  -cnv -p "Write a robust retry helper in Python."

Exact file sizes, SHA-256 hashes, imatrix settings, and smoke-test throughput are recorded in release-manifest.json.