Reflex-jr/Qwen3.6-35B-A3B-Domain-Aggressive-GGUF

Qwen3.6-35B-A3B-Domain-Aggressive — GGUF ladder

GGUF quantizations of `Reflex-jr/Qwen3.6-35B-A3B-Domain-Aggressive` — a domain-fine-tuned, refusal-direction-orthogonalized variant of Qwen/Qwen3.6-35B-A3B.

Built for the AMD Developer Hackathon 2026 by @Jayanth-reflex — solo dev in India, 48-hour sprint on a single AMD Instinct MI300X (192 GB HBM3, gfx942).

Source repo: https://github.com/Jayanth-reflex/amd-hackathon-2026

What this is

A 35B-parameter MoE language model (256 experts, top-8 routing) fine-tuned with LoRA on an Indian-context domain blend (law, taxation, pharmacology, advanced game theory, engineering, geopolitics, psychology, survival, languages including Telugu and Hindi), then weight-orthogonalized via the Arditi/Labonne method (5 directions, layers 36→32) to reduce its baseline refusal posture.

Architecture: Qwen3.6 hybrid attention (10 full self-attention layers + 30 linear-attention "chunkgateddelta_rule" layers) + 256-expert MoE per layer. 35B total / ~3B active per token.

Active parameters: ~3B (MoE routing) — so quantized inference latency is closer to 3B-class than 35B-class, while quality stays at 35B-class.

Quant ladder

-imatrix.gguf is the calibration file used to derive the importance-weighted quantizations (calibrated on bartowski's calibration_datav3 corpus, 280 KB / 66k tokens of mixed text).

Usage

llama.cpp (recommended for portability)

# Build llama.cpp with HIP (AMD)
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
  cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx942 \
  -DGGML_HIP_ROCWMMA_FATTN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

# Or build for CUDA / Metal / CPU as usual

# Serve as OpenAI-compatible API
./build/bin/llama-server \
  -m Qwen3.6-35B-A3B-Domain-Aggressive-Q4_K_M.gguf \
  --jinja -c 32768 -ngl 99 -fa \
  -ctk bf16 -ctv bf16 \
  --parallel 4 \
  --host 0.0.0.0 --port 8001

Key flag: Use -ctk bf16 -ctv bf16 (NOT default f16). Qwen3.6 is trained in bf16 and f16 KV measurably degrades perplexity.

Apple Silicon failover (M-series Mac, 24+ GB unified memory)

./build/bin/llama-server \
  -m Qwen3.6-35B-A3B-Domain-Aggressive-Q4_K_M.gguf \
  --jinja -c 32768 -ngl 99 \
  --host 0.0.0.0 --port 8001

Q4KM comfortably fits in 24 GB unified memory with room for context.

Ollama

ollama create qwen3.6-domain -f Modelfile
# where Modelfile points at the Q4_K_M .gguf

Sampling recommendations

temperature: 0.7
top_p: 0.95
presence_penalty: 1.5  # Qwen-family default

For domain Q&A (law, tax, etc.), drop temperature to 0.3 for more deterministic output. For creative writing, raise to 0.9.

Safety / policy posture

This model is abliterated — its baseline refusal posture has been reduced via Arditi/Labonne weight orthogonalization (5 directions, 605 weight-tensor updates across embed + 10 selfattn + 30 linearattn + 40 shared_expert + 40 fused MoE-expert tensors). Refusals are still present but in softer, less moralistic language.

Policy enforcement is intended to happen at the application layer, not in the weights. The reference implementation pairs this model with meta-llama/Llama-Guard-3-1B at the gateway, with hard-line refusals for: CSAM (S4), bioweapon synthesis (S9), narrow controlled-substance synthesis (S2), targeted impersonation (S5), doxxing (S7).

If you operate this model without a classifier-tier safety wrapper, you assume responsibility for what it outputs. The point of the architectural separation is so policy lives in inspectable code (policy.yaml), not in opaque baked-in alignment.

See https://github.com/Jayanth-reflex/amd-hackathon-2026/blob/main/docs/SAFETY.md for the full reference architecture.

License chain

• —Base model Qwen/Qwen3.6-35B-A3B: Apache-2.0
• —Our fine-tune (LoRA → merged): Apache-2.0 (downstream of base)
• —Our wrapper code (training scripts, gateway, frontend): MIT (in the public repo)
• —Abliteration was implemented in our own code (no Heretic dependency); MIT.

Citation

@misc{reflexjr-2026-qwen36-domain,
  author       = {Jayanth (Reflex-jr)},
  title        = {Qwen3.6-35B-A3B-Domain-Aggressive (GGUF)},
  year         = 2026,
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/Reflex-jr/Qwen3.6-35B-A3B-Domain-Aggressive-GGUF},
  note         = {AMD Developer Hackathon 2026 submission}
}

Acknowledgements

• —Maxime Labonne (blog) for the canonical practical guide to abliteration
• —Arditi et al. 2024 (paper) for the original method
• —bartowski for calibration_datav3 (the imatrix corpus de-facto standard)
• —The Qwen team at Alibaba for the base model
• —The llama.cpp / GGML team for the inference stack
• —AMD for the MI300X compute grant + the Instinct gfx942 hardware