README.md
| 1 | --- |
| 2 | license: apache-2.0 |
| 3 | pipeline_tag: text-generation |
| 4 | library_name: llama.cpp |
| 5 | tags: |
| 6 | - gguf |
| 7 | - qwen3_5 |
| 8 | - image-text-to-text |
| 9 | - agentic |
| 10 | - reinforcement-learning |
| 11 | - tool-use |
| 12 | - hybrid-attention |
| 13 | - qwen3.8 |
| 14 | - conversational |
| 15 | base_model: |
| 16 | - sabicap/sabi-noema |
| 17 | --- |
| 18 | |
| 19 | ## Sabi-Noema GGUF Quantizations |
| 20 | |
| 21 | See the [original model card](https://huggingface.co/sabicap/sabi-noema) for details (description, benchmarks, generation parameters). |
| 22 | |
| 23 | ## Speculative decoding |
| 24 | |
| 25 | sabi didn't include spec decoding tensors, but you can try one from any other Qwen3.8-27B GGUFs. For example: |
| 26 | - [ggml-org/Qwen3.8-27B-GGUF](https://huggingface.co/ggml-org/Qwen3.8-27B-GGUF/tree/main) — MTP and DFLASH |
| 27 | - [unsloth/Qwen3.8-27B-GGUF](https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/tree/main/MTP) — MTP only |
| 28 | |
| 29 | I'm getting around 40%+ acc rate with Unsloth Q4_0 MTP. |
| 30 | |
| 31 | ## Usage |
| 32 | |
| 33 | ```bash |
| 34 | llama-cli -m Sabi-Noema-Q4_K_M.gguf --mmproj mmproj-Sabi-Noema-BF16.gguf --model-draft mtp-Qwen3.8-27B-Q4_0.gguf --spec-type draft-mtp |
| 35 | ``` |
| 36 | |
| 37 | ## Conversion |
| 38 | |
| 39 | llama.cpp [`23deec025`](https://github.com/ggml-org/llama.cpp/commit/23deec025) (v0.6.0) |
| 40 | |