| base_model: | |
| - llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF | |
| library_name: llama.cpp | |
| license: apache-2.0 | |
| tags: | |
| - gguf | |
| - qwen3.6 | |
| - qwen3_5_moe | |
| - moe | |
| - mtp | |
| - uncensored | |
| - heretic | |
| - alice | |
| - chinese | |
| - roleplay | |
| # Alice Qwen3.6 35B A3B MTP GGUF | |
| Alice shell release for `Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q4_K_M.gguf`. | |
| This release patches only `tokenizer.chat_template` to add a shallow Alice identity and roleplay-oriented default behavior. Model tensors are unchanged. | |
| Base repo: `llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-GGUF` | |
| Recommended llama.cpp flags for MTP smoke testing: | |
| ```bash | |
| llama-server \ | |
| -m Alice-Qwen3.6-35B-A3B-MTP-Q4_K_M.gguf \ | |
| -c 8192 -ngl 99 -np 1 \ | |
| --spec-type draft-mtp --spec-draft-n-max 2 \ | |
| --reasoning off --jinja | |
| ``` | |
| Smoke result: | |
| - MTP initialized with `draft-mtp`. | |
| - Identity prompts anchor as Alice. | |
| - Companion prompt generated normally. | |
| - Example MTP acceptance: `41 accepted / 114 generated` draft tokens on a short Chinese companion prompt. | |
| Notes: | |
| - Keep `np=1` for first tests. | |
| - Test text-only first; do not attach `mmproj` until text generation is stable. | |
| - If the runtime does not expose MTP, the model should still run as a normal GGUF. | |