A 4-bit MLX conversion of the Qwen3.5-122B-A10B-Opus-Reasoning fine-tune, built to run fast on Apple Silicon. Converted from the BF16 source and quantized to 4-bit (group size 64, affine) with router/gate tensors kept at 8-bit for stability. This model should run great on machines with either 96GB (what I tested) or 128GB of memory.

I built this quant to work well with my local AI workspace tool, Familiar. Familiar is an AI-enabled workspace where you can take Notes, collaborate in wikis, chat with models, and create smart automations with local models. This model performs well for most activities that Familiar needs. Check it out here:

https://github.com/sixvolts/familiar

Highlights

Architecture: qwen3_5_moe — a hybrid linear-attention (Gated DeltaNet) + full-attention MoE (full_attention_interval=4)
Size: 122B total params, ~10B active per token (A10B)
Experts: 256 routed, 8 active per token + shared expert
Layers: 48 · hidden: 3072 · vocab: 248320
Quantization: 4-bit, group size 64, affine (~4.5 bits/weight); MoE router gate and shared_expert_gate at 8-bit
On disk: ~64 GB
Speed: ~60 tok/s generation on an M3 Ultra (60-core GPU)

Use with MLX

pip install mlx-lm

bash

# text generation (text-only loader)
python -m mlx_lm generate \
  --model SixVolts/Qwen3.5-122B-A10B-Opus-Reasoning-MLX \
  --prompt "Explain how a bicycle stays upright." \
  --max-tokens 1024

python

from mlx_lm import load, generate

model, tokenizer = load("SixVolts/Qwen3.5-122B-A10B-Opus-Reasoning-MLX")
messages = [{"role": "user", "content": "Show your steps: what is 240 / 150 in hours and minutes?"}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=1024))

Serving an OpenAI-compatible endpoint

bash

python -m mlx_lm server \
  --model SixVolts/Qwen3.5-122B-A10B-Opus-Reasoning-MLX \
  --host 127.0.0.1 --port 8081

Note: As a reasoning model, the thinking trace is returned in the response's reasoning field and the final answer in content. Allow a generous max_tokens (2000+) so generation has room to finish thinking and answer — otherwise content can come back empty because the budget was spent reasoning.

Notes & requirements

Built and validated with mlx-lm. Use the text-only loader (mlx_lm), not mlx_vlm.
Runs comfortably on a Mac with ~96 GB unified memory (peak ~69 GB at this quant).
Recommended sampling (from generation_config.json): temperature=0.6, top_p=0.95, top_k=20.

Provenance

Converted from the BF16 weights of timteh673/Qwen3.5-122B-A10B-Opus-Reasoning-GGUF to Hugging Face safetensors, then quantized to 4-bit with mlx-lm. The GGUF→safetensors step reproduces the upstream Qwen3.5 tensor transforms (Gated-DeltaNet v-head reordering and the A_log = log(-ssm_a) SSM-decay transform) so long-form generation stays coherent.