Qwen3.6-35B-A3B MTPLX Optimized Balance
Balanced local 35B-A3B inference for Apple Silicon, packaged for MTPLX native Multi-Token-Prediction speculative decoding.
This checkpoint is the balanced 35B release: a 6-bit MLX body paired with calibrated INT4 MTP heads. It is tuned for strong reasoning-on generation speed while keeping prompt processing and memory use practical across normal coding contexts.
Run It
brew install youssofal/mtplx/mtplx
mtplx start
mtplx run "hello" --model Youssofal/Qwen3.6-35B-A3B-MTPLX-Optimized-Balance
For an OpenAI-compatible local server:
mtplx serve --model Youssofal/Qwen3.6-35B-A3B-MTPLX-Optimized-Balance --profile sustained --max --port 8000 --no-stats-footer
Why This Exists
MTPLX uses the model's own MTP heads to generate draft tokens, then verifies them with the main model. When the draft heads are well-matched, you get higher throughput without running a separate drafter model.
Optimized Balance is built for that path. MTPLX reads mtplx_runtime.json and
selects the measured D2 defaults automatically.
Recommended Runtime Defaults
| Setting | Value |
|---|---|
| Backend | qwen3-next-mtp |
| Default depth | D2 |
| Target sampler | temp=0.60, top_p=0.95, top_k=20 |
| Draft sampler | temp=0.60, top_p=0.95, top_k=20 |
| Profile | sustained |
| Benchmark fan mode | max |
Performance
Measured in MTPLX Sustained Max on Apple Silicon with reasoning enabled.
Generation
| Mode | TPS | Verify time | Acceptance |
|---|---|---|---|
| AR baseline | 86.30 | - | - |
| D1 comparison | 123.00 | 7.24s | 0.8329 |
| D2 promoted default | 126.43 | 6.62s | 0.8134, 0.5048 |
| D3 comparison | 112.43 | 7.16s | 0.7802, 0.4709, 0.2514 |
D2 is the promoted default because it gives the best balance of throughput, acceptance, and verify cost. A three-run D2 repeat averaged 123.44 tok/s, with every run above 122 tok/s.
Prompt Processing
| Context | Prompt TPS |
|---|---|
| 512 | 1,756.8 |
| 1k | 3,339.2 |
| 2k | 4,109.8 |
| 4k | 4,048.7 |
| 8k | 3,872.0 |
| 16k | 3,162.3 |
| 32k | 2,761.3 |
| 64k | 1,834.2 |
Average prompt processing across the 512-to-64k ladder was 3,110.5 tok/s.
Model Build
| Component | Format |
|---|---|
| Main body | 6-bit MLX affine, group size 64 |
| Router and gate tensors | 8-bit where recorded by config |
| MTP numbered-expert weights | INT4 MLX affine, group size 64 |
| Norms, scales, biases, plain tensors | BF16 |
This is not a full-precision checkpoint. It is built for fast local use on Apple Silicon through MTPLX.
Files
model-*.safetensors: MLX 6-bit body shardsmtp.safetensors: calibrated INT4 MTP sidecarmtplx_runtime.json: MTPLX runtime contract and measured defaultsMTPLX_PUBLISH_MANIFEST.json: file sizes and benchmark summary- tokenizer and config files for local loading