Qwen3.6-35B-A3B Claude 4.7 Opus Reasoning Distilled - MLX oQ4 MTP

MLX/oMLX 4-bit conversion of r3lax/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled with Qwen MTP tensors preserved and runtime-tested in oMLX.

This is not a new training run or fine-tune. Weights were only converted/quantized for local MLX/oMLX inference.

Quick Facts

• Architecture: Qwen3.6 35B-A3B MoE, roughly 3B active parameters per token.
• Quantization: oQ4-style MLX 4-bit, group size 64.
• MTP: preserved and verified in oMLX.
• Test hardware: Apple Silicon M5 Pro with 48GB unified memory.
• Runtime tested with oMLX native MTP enabled.

MTP Verification

MTP config: mtp_num_hidden_layers=1, mtp_use_dedicated_embeddings=false
MTP tensor entries: 42
MTP fusion projection: language_model.mtp.fc.weight present and full precision
Runtime: oMLX logged native MTP path activation during smoke tests

MTP support is runtime-specific. These tensors are preserved, but non-oMLX runtimes may ignore them.

Local Speed Smoke Test

Measured on an M5 Pro Mac with 48GB unified memory using oMLX.

MTP on average:  ~89.9 tok/s
MTP off average: ~86.0 tok/s
Speed lift:       ~+4.6% with native MTP enabled

Per-prompt MTP-on smoke results:

count120:   160 tokens, 1.726s, 92.69 tok/s, MTP accept 79/79 and 79/79
rain180:    220 tokens, 2.465s, 89.26 tok/s, MTP accept 96/123 and 98/121
jsoncities: 128 tokens, 1.457s, 87.86 tok/s, MTP accept 64/65 and 64/64

These are local smoke numbers, not a universal benchmark. Prompt, cache state, batching, oMLX version, and hardware will change results.

Usage Notes

Place the model under your oMLX model directory and enable native MTP:

~/.omlx/models/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-MLX-oQ4-MTP

Recommended oMLX settings:

{
  "mtp_enabled": true,
  "dflash_enabled": false,
  "turboquant_kv_enabled": false
}

Attribution

Credit to the upstream work:

• Qwen/Qwen3.6-35B-A3B for the base model.
• lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled for the Claude Opus 4.7 reasoning distillation.
• r3lax/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled for the source checkpoint used here.
• Anthropic Claude Opus 4.7 as the teacher model used by the upstream distillation.
• oMLX / MLX community for the Apple Silicon runtime and MTP support.

Please credit the upstream authors if you use or redistribute this conversion.

Caveats

• MTP runtime activation was verified in oMLX only.
• This conversion is experimental.
• Reasoning models can emit long <think> traces; set max_tokens intentionally.
• The upstream model card notes that distillation transfers reasoning style, not new factual knowledge.

License

Apache-2.0, following the upstream model metadata. Check upstream model cards and any teacher-data usage policies before redistribution or commercial deployment.