Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-Text-MLX-oQ8

Mixed-precision quantization for Apple Silicon, text-only mode (vision tower stripped). Highest quality text-only version — closest to BF16 baseline.

Quantized from lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled using oMLX's oQ8 algorithm (sensitivity-aware mixed-precision quantization).

📊 Specs

Field	Value
Base model	Qwen/Qwen3.6-35B-A3B (35B params, 128 experts MoE, A3B activation)
Fine-tune	LoRA distilled from Claude 4.7 Opus reasoning outputs (lordx64 dataset)
Quantization	oMLX oQ8 (mixed-precision, ~8.7 bpw average)
Modality	Text only (vision tower stripped)
Format	MLX safetensors
Model size	~34 GB
Inference memory	~37 GB (incl. KV cache and runtime overhead)
Recommended hardware	Apple Silicon M2 Ultra 64GB+ / M3 Max / M5 Max

🚀 Quick Start

Install

pip install mlx-lm
# Or with uv:
uv tool install mlx-lm

Inference

mlx_lm.generate \
  --model wangkezun/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-Text-MLX-oQ8 \
  --prompt "Explain mixture of experts in one paragraph." \
  --max-tokens 512

Python API

from mlx_lm import load, generate

model, tokenizer = load(
    "wangkezun/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-Text-MLX-oQ8"
)

response = generate(
    model,
    tokenizer,
    prompt="Solve: integrate x*sin(x) dx",
    max_tokens=512,
)
print(response)

OpenAI-compatible Server

mlx_lm.server \
  --model wangkezun/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-Text-MLX-oQ8 \
  --port 8080

Drop-in compatible with OpenAI clients including Claude Code (with custom backend), AstrBot, Open WebUI, LibreChat, and Continue.dev.

📈 Measured Performance

Benchmarked on MacBook Pro M5 Max 128GB:

Metric	Value
Prompt processing	~58 tokens/s
Generation speed	~103 tokens/s
Peak memory	36.9 GB
Model load time	~12 sec

When to Choose oQ8 over oQ6

oQ8 retains slightly more precision than oQ6, but the observable quality difference is small for most tasks. Use oQ8 when:

• You need the absolute highest fidelity quantization
• Running quality benchmarks against the BF16 reference
• Memory budget is generous (37+ GB free)

For most users, Text-oQ6 is the better choice — comparable quality at 7 GB less footprint.

Why Text-Only?

Stripping the vision tower offers practical advantages for text-only workflows:

• ~10% faster generation vs the VLM equivalent (no vision compute path overhead)
• ~2 GB less peak memory
• Simpler deployment — no need for vision processor configs or image preprocessing dependencies

If you only feed text into your model, this version is strictly better than the VLM variant.

🧠 Model Behavior

Inherits the Claude reasoning distillation: the model uses <think>...</think> tags to structure its chain-of-thought before producing the final response.

Best for:

• Coding agents and tool-use workflows
• Complex reasoning tasks (math, logic, analysis)
• Quality-sensitive applications where size is not a constraint
• Quantization quality reference / baseline comparisons

Sample output structure:

<think>
1. Analyze the user's request: ...
2. Identify the key constraints: ...
3. Formulate the solution: ...
</think>

Here is my analysis: ...

🔬 Quantization Details

• Source model: Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled (BF16 MLX-converted)
• Sensitivity model: 8-bit quantization of the same distilled model (self-referenced sens for tight distribution alignment)
• Non-quant weight dtype: bfloat16 (M3+ optimal)
• Text-Only mode: ON (vision tower stripped — verified: 0 vision-related tensors)
• Quantizer: oMLX

📦 Other Versions in This Series

Version	Size	Best for
VLM-MLX-oQ4	19.6 GB	Memory-constrained inference (with vision)
VLM-MLX-oQ6	27 GB	Recommended VLM quality/size ratio
VLM-MLX-oQ8	35 GB	VLM quality reference baseline
Text-MLX-oQ4	19 GB	Text-only, fastest
Text-MLX-oQ6	27 GB	Recommended text-only
Text-MLX-oQ8	34 GB	Text-only, max quality

Choosing a version:

• Text-only workflows (coding, agents, dialogue) → Text variants are faster and lighter
• Image input needed (OCR, visual analysis, screenshot understanding) → VLM variants
• oQ6 is the sweet spot for most use cases. oQ8 yields diminishing returns relative to its size.

⚠️ Disclaimer

This model derives from a chain of upstream work:

1. Base model Qwen/Qwen3.6-35B-A3B by Alibaba's Qwen team (Apache-2.0)
2. Distilled variant lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled by lordx64 using Claude 4.7 Opus reasoning outputs (Apache-2.0)
3. This quantization by @wangkezun using oMLX on Apple Silicon

This model is not affiliated with or endorsed by Anthropic, PBC. "Claude" is a trademark of Anthropic, PBC. The use of "Claude" in this model name is purely descriptive (nominative fair use) to indicate the upstream training data lineage.

By using this model, you agree to comply with:

• The Apache-2.0 license inherited from the base model
• Any applicable license terms of the upstream distillation dataset
• Local laws and regulations governing AI model usage in your jurisdiction

🙏 Acknowledgments

• Alibaba Qwen Team — for the Qwen3.6-35B-A3B base model
• lordx64 — for the reasoning-focused LoRA distillation
• Jundot (oMLX team) — for the oQ mixed-precision quantization algorithm
• Apple MLX team — for the MLX framework and tooling

📜 License

Apache-2.0 (inherited from base model).

---

Generated: 2026-04-26
Quantizer: @wangkezun