MTPLX-branded multi-token-prediction model optimized for Apple Silicon (MLX).
Forged with MTPLX Forge from mlx-community/Qwythos-9B-v2-OptiQ-4bit.

Verification & Benchmarks

Metric	Result
Speedup	1.85× vs. autoregressive baseline
Throughput	67.1 tok/s (Baseline: 36.3 tok/s)
Best Depth	`D2`
Mean Acceptance Rate	93% at `D2`
Hardware / OS	Apple M4 Pro · macOS 26.6.2
Sampler Config	`temp=0.6`, `top_p=0.95`, `top_k=20`

Refer to mtplx_runtime.json for the complete verification log.

Usage

MTPLX detects and initializes the model automatically once downloaded:

bash

# Pull model weights
mtplx pull nRanzo/mlx-community-Qwythos-9B-v2-OptiQ-4bit-MTPLX

# Launch interactive session
mtplx start chat

Project & Implementation

For benchmarks, speculative decoding pipelines, and testing methodology, visit the MLX-MTP Repository.

Attribution & License

Base Model: mlx-community/Qwythos-9B-v2-OptiQ-4bit (MLX Community)
License: Apache-2.0 (LICENSE)