Add Alice Qwen3.6 27B Code Heretic MTP Q8_0 GGUF

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+Alice-Qwen3.6-27B-Code-Heretic-MTP-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text

Alice-Qwen3.6-27B-Code-Heretic-MTP-Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2588f4ac7aca14376ae01d02623c983925f3faee54276d189027faf2f931ab76
+size 29047084000

README.md

@@ -0,0 +1,71 @@
+# Alice-Qwen3.6-27B-Code-Heretic-MTP-GGUF
+
+High-fidelity Q8_0 GGUF build for Alice's 27B code/security lane.
+
+## Files
+
+- `Alice-Qwen3.6-27B-Code-Heretic-MTP-Q8_0.gguf`
+
+## Base
+
+- Source base: `huginnfork/Qwen3.6-27B-uncensored-heretic-v2-mtp`
+- Architecture: Qwen3.5 / Qwen3.6 hybrid, GGUF architecture key `qwen35`
+- License: Apache-2.0 from the source base
+- This package does not train or alter model tensors. It preserves the uncensored/heretic base and adds an Alice code/security chat template in GGUF metadata.
+
+## Alice Profile
+
+The bundled chat template anchors the model as Alice for the current conversation and biases the model toward:
+
+- code generation
+- Solidity and smart contract work
+- security review and engineering fixes
+- direct Chinese/English technical answers
+- honest reporting when code, logs, files, or tool outputs are missing
+
+This is the code/security line, not the RP/NSFW line.
+
+## Runtime Notes
+
+This Q8_0 build is for higher-end local runners. The GGUF file is about 27 GiB on disk and needs additional memory for KV cache and runtime overhead.
+
+Recommended targets:
+
+- Mac with 64 GB+ unified memory, 96 GB preferred for larger context
+- Linux/Windows rigs with 40 GB+ practical available VRAM/RAM, depending on offload and context
+- 24 GB GPUs should use lower quantization unless context is kept small and CPU offload is acceptable
+
+MTP speculative decoding is bundled in this GGUF. A local `llama-cli` smoke test with `--spec-type draft-mtp --spec-draft-n-max 2` ran successfully.
+
+## Local Smoke Results
+
+Environment:
+
+- Host: Apple Silicon Mac, local llama.cpp Metal build
+- Context: `-c 4096`
+- GPU offload: `-ngl 99`
+- Sampling: `temp=0.2`, `top_p=0.8`
+
+Observed short-test speed:
+
+- Normal generation: about 11 tok/s
+- MTP generation: about 19 tok/s on a short identity prompt
+
+Smoke prompts passed:
+
+- identity: answers as Alice, no thinking leak
+- Solidity Counter: emits SPDX, pragma, `contract Counter`, `uint256 public count`, `increment()`
+- Reentrancy repair: identifies CEI fix and moves balance update before external call
+
+## Example llama.cpp Command
+
+```bash
+./llama-cli \
+  -m Alice-Qwen3.6-27B-Code-Heretic-MTP-Q8_0.gguf \
+  -ngl 99 -c 4096 \
+  --spec-type draft-mtp --spec-draft-n-max 2 \
+  --reasoning off \
+  -cnv -st \
+  -p "你是谁？然后用一句话说你擅长什么。"
+```
+