Hemmingway-1 heretic GGUF
arelath/Hemmingway-1-heretic-adapter merged into Altworld/Hemmingway-1 in BF16 with PEFT, then converted and quantized with llama.cpp. Exact source commits are in provenance.json.
The original base model's multi-token prediction (MTP) head is included in each GGUF. The adapter modifies the main model; the MTP head is preserved unchanged before quantization. In a compatible llama.cpp runtime, enable it with --spec-type draft-mtp. Draft acceptance and speedup with the modified model have not been benchmarked.
Bit levels are nominal quantization families; effective storage per parameter is higher. No importance matrix was used.
| Nominal bits | File | Preset |
|---|---|---|
| 8 | Hemmingway-1-heretic-MTP-Q8_0.gguf | Standard Q8_0 |
| 6 | Hemmingway-1-heretic-MTP-Q6_K.gguf | Standard Q6_K |
| 5 | Hemmingway-1-heretic-MTP-Q5_K_S.gguf | Smaller 5-bit K-quant mix |
| 5 | Hemmingway-1-heretic-MTP-Q5_K_M.gguf | Medium 5-bit K-quant mix |
| 4 | Hemmingway-1-heretic-MTP-Q4_K_S.gguf | Smaller 4-bit K-quant mix |
| 4 | Hemmingway-1-heretic-MTP-Q4_K_M.gguf | Medium 4-bit K-quant mix |
| 3 | Hemmingway-1-heretic-MTP-Q3_K_S.gguf | Smaller 3-bit K-quant mix |
| 3 | Hemmingway-1-heretic-MTP-Q3_K_M.gguf | Medium 3-bit K-quant mix |
| 2 | Hemmingway-1-heretic-MTP-Q2_K.gguf | Standard 2-bit K-quant mix |
| 2 | Hemmingway-1-heretic-MTP-TQ2_0.gguf | 2-bit ternary quantization (experimental quality) |
Use a recent llama.cpp-compatible runtime supporting Qwen3.5. No quality or inference evaluation has been performed by this pipeline.