Running Qwen3.6 Locally: Abliterated Variant

Works with v2.4+ builds that include the llama.cpp GGUF engine (opens in a new tab); see the base recipe (opens in a new tab).

This directory is the Qwen3.6 recipe (opens in a new tab) with the model swapped for huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated (opens in a new tab), huihui-ai's own Q4_K GGUF (21.7 GB) of Qwen3.6-35B-A3B with its refusal behavior removed by abliteration. The file keeps the model's multi-token-prediction (NextN) layer, so Spice drafts with it automatically. It serves the model and its API; it does not load the dataset or enable tools, for the reason in Why no dataset tools (opens in a new tab).

Abliteration removes the "refusal direction" from a model's weights, so the model answers requests the base model would decline. huihui-ai describes its method as a proof-of-concept implementation of refusal removal (opens in a new tab). The result has none of the base model's safety training. Use it for research, red-teaming and controlled experiments, add your own moderation before exposing it to users, and review its outputs. huihui-ai's model card carries its own usage warnings and disclaims responsibility for use.

Spicepod

version: v2
kind: Spicepod
name: qwen3.6-abliterated

models:
  - from: huggingface:huggingface.co/huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF
    name: qwen3.6-abliterated
    files:
      - path: Huihui-Qwen3.6-35B-A3B-abliterated-ggml-model-Q4_K.gguf

version: v2
kind: Spicepod
name: qwen3.6-abliterated

models:
  - from: huggingface:huggingface.co/huihui-ai/Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF
    name: qwen3.6-abliterated
    files:
      - path: Huihui-Qwen3.6-35B-A3B-abliterated-ggml-model-Q4_K.gguf

Run

cd models/qwen3.6/abliterated
spice run

cd models/qwen3.6/abliterated
spice run

Wait for Model [qwen3.6-abliterated] deployed, ready for inferencing, then in a second terminal:

spice chat

spice chat

The API serves the model as qwen3.6-abliterated at http://localhost:8090/v1/chat/completions, with the same reasoning_effort behavior as the base model (none for a direct answer).

Why no dataset tools

This GGUF carries Qwen's stock chat template, which renders a tool call's arguments with tool_call.arguments|items. When tools: auto is set, Spice primes the conversation with a list_datasets call whose arguments are an empty string, and the stock template fails to render it: every request returns HTTP 400 Failed to render the chat template: invalid operation: cannot convert value into pairs. Unsloth's templates guard that loop, which is why the base recipe's tools work. A chat_template override does not help, because a GGUF's own template takes precedence. The runtime bug is spiceai/spiceai#15578 (opens in a new tab); once it is fixed, add the base recipe's datasets block and params: { tools: auto } to this spicepod.

Memory

The repository holds Q2_K (13.2 GB, 16 GB Macs), Q3_K (17.2 GB, 24 GB to 32 GB), Q4_K (21.7 GB, the default, 36 GB and up), Q5_K (25.3 GB), Q6_K (29.2 GB, 48 GB and up) and Q8_0 (37.8 GB, 64 GB and up), all with the NextN layer. For the dense 27B model, huihui-ai/Huihui-Qwen3.6-27B-abliterated-MTP-GGUF (opens in a new tab) has the same ladder (Q4_K, 16.8 GB).

Clean up

rm -rf ~/.cache/huggingface/hub/models--huihui-ai--Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF

rm -rf ~/.cache/huggingface/hub/models--huihui-ai--Huihui-Qwen3.6-35B-A3B-abliterated-MTP-GGUF