license: apache-2.0 license_link: https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/LICENSE base_model:

• llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved
• Qwen/Qwen3.6-35B-A3B tags:
• gguf
• qwen3.6
• qwen3_5_moe
• qwen35moe
• moe
• text-generation
• conversational
• uncensored
• heretic
• abliterated
• decensored
• mtp
• q6dense
• gufo
• strix-halo pipeline_tag: text-generation library_name: gguf language:
• en

Qwen3.6-35B-A3B Uncensored Heretic — Native MTP — Q6dense

Q6dense GGUF quantization of llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved, preserving the native MTP block.

This is a standard GGUF. “Q6dense” here refers to the per-tensor quantization recipe (dense trunk at Q6_K, routed experts at Q4_K/Q5_K, embeddings and MTP dense weights at Q8_0), not a new file format. It is tuned for fast inference on AMD Strix Halo / gfx1151 with gufo, but it also loads in recent llama.cpp builds that support qwen35moe and native MTP.

⚠️ Content warning: This is an uncensored/“heretic” fine-tune. It can produce offensive, explicit, unsafe, or factually wrong output. There is no safety filtering. You are responsible for how you use it and for complying with applicable laws and platform policies.

Model details

Files

The Q6dense file alone is text-only. Image/video input needs the separate mmproj file and a runtime that supports it.

Quantization recipe

Produced from the BF16 GGUF with llama.cpp b11069:

llama-quantize \
  --imatrix imatrix_unsloth.gguf_file \
  --tensor-type-file q6dense.types \
  Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-BF16.gguf \
  Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q6dense.gguf \
  Q4_K_M 16

Per-tensor types:

• Dense trunk attention, SSM, shared-expert and output tensors: Q6_K
• Routed experts (ffn_*_exps): mostly Q4_K for gate/up, Q5_K for down
• Token embeddings and MTP dense tensors: Q8_0
• Norms and small SSM parameters: F32

This is a lossy quantization. It is not bit-identical to the BF16 or Q8_0 release.

Usage

gufo (Strix Halo / gfx1151)

gufo serve llm \
  --host 0.0.0.0 --port 8080 \
  --model Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q6dense.gguf \
  --speculative mtp \
  --sessions 1 \
  --context 65536 \
  --think on \
  --reasoning-effort medium

• --speculative mtp uses the native MTP block in this GGUF. No sidecar/draft model is needed.
• --think on|off controls the thinking phase. --reasoning-effort accepts auto, minimal, low, medium, high, xhigh, max.
• Keep --context at a value you can afford: roughly 34 KiB/token with MTP. 32K ≈ 1.1 GiB KV, 64K ≈ 2.1 GiB, 256K ≈ 8.5 GiB.

Chat-template note for gufo: this fine-tune carries its own Qwen3.6 template with SHA-256 e7018ebe03fa07c44c7eea46fa3734173a90237a02b7c352d3260914b40d7c6f. Some older gufo builds only accept the official and Unsloth template hashes. If you get Unsupported Qwen3.6 chat template, add the hash above to IsQwen36Template() in src/models/qwen/chat_template.cpp, or use a gufo build that accepts it.

llama.cpp

Use a recent build with qwen35moe and MTP support:

llama-server \
  -m Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved-Q6dense.gguf \
  --jinja \
  --ctx-size 32768 \
  -fa on \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --spec-type draft-mtp --spec-draft-n-max 5

For Open WebUI, point the OpenAI-compatible connection at http://127.0.0.1:8080/v1.

Prompt format and sampling

The GGUF includes the fine-tune's chat template. It defaults to thinking on.

To disable thinking in an OpenAI-compatible request:

{
  "chat_template_kwargs": {
    "enable_thinking": false
  }
}

System prompts are supported. This fine-tune's template also accepts developer messages and <|think_on|> / <|think_off|> markers, but support for those depends on the runtime. gufo uses its own compiled formatter and ignores those template extensions.

Memory guidance

Approximate GPU/unified-memory use:

On a 128 GiB Strix Halo system this leaves enough room to run an image-generation stack alongside it. Reduce --context and --sessions if you run other GPU workloads.

Limitations

• Uncensored fine-tune. No safety alignment or filtering is guaranteed. Outputs may be harmful, biased, or illegal in your jurisdiction.
• Quantized. Q6dense introduces quantization loss versus BF16/Q8_0. Expect small quality and reasoning differences.
• No formal evaluation. No benchmark or safety evaluation was performed specifically for this quantized artifact. Results may differ from the base model and from the original fine-tune.
• Text-only here. Vision requires the separate mmproj file and runtime support.
• Template compatibility. The embedded chat template is not the official Qwen3.6 template hash, so some runtimes with strict template pinning may reject it until whitelisted.
• Base-model limitations. Inherits the knowledge cutoff, biases, and failure modes of Qwen3.6-35B-A3B and the fine-tune.

Credits

• Base model: Qwen/Qwen3.6-35B-A3B
• Uncensored/heretic fine-tune: llmfan46
• Q6dense recipe and model support: gufo
• q6dense.types: ligamentexceed/Qwen3.6-35B-A3B-Q6dense-GGUF
• Imatrix: unsloth/Qwen3.6-35B-A3B-MTP-GGUF
• Quantization tool: llama.cpp

License

Apache-2.0, inherited from Qwen3.6-35B-A3B. The fine-tune and this quantization are redistributed under the same terms. See the upstream repositories for full license and use restrictions.

Quantized by Yanuk-K on 10/19/26. BF16 source SHA-256: e7018ebe03fa07c44c7eea46fa3734173a90237a02b7c352d3260914b40d7c6f. Q6dense file SHA-256: e7018ebe03fa07c44c7eea46fa3734173a90237a02b7c352d3260914b40d7c6f.