MTPLX V2 hits 82 TPS on Qwen 3.6 27b with Turbo Mode
New custom quantized-matmul kernels deliver 82 tokens/sec on a MacBook Pro M5 Max
Deep Dive
MTPLX V2, released one month after V1, introduces Turbo Mode using custom verify-specialized quantized-matmul kernels and a compiled verify step, achieving 82 TPS on a MacBook Pro M5 Max at temperature 0.6. It also includes significant changes to SSD KV cache and long context tool calling improvements. Preliminary benchmarks from Ivan Fioravanti compare MTPLX against oMLX and DGX spark.
Key Points
- 82 tokens per second achieved on Qwen 3.6 27b using a MacBook Pro M5 Max at temperature 0.6
- Turbo Mode uses custom verify-specialized quantized-matmul kernels with a compiled verification step
- Includes SSD KV cache improvements and enhanced long context tool calling support
Why It Matters
On-device AI execution reaches laptop-native speeds, enabling responsive local coding assistants and large model inference without cloud costs.