MTPLX 2.10 vs oMLX: Qwen3.8-Flash-Next coding tests

The speaker compares MTPLX v2.10.0 with oMLX for running Qwen3.8-Flash-Next as a local LLM with a coding assistant. His test machine is an Apple M5 Max with 128 GB RAM, and he uses a 4-bit model. He reports chat speeds near 70 tokens per second and roughly 50 to 60 during agent use; these figures describe his setup rather than a general performance guarantee.

The release overview covers draft tokens and multi-token prediction, NVMe SSD offloading with N-gram support, long-context optimization, and memory-aware KV caching. The speaker attributes support for more than 147,000 context tokens and controlled SSD spill to the project's author. He also describes an 8-bit option aimed at output quality, though the demonstrations use 4-bit quantization.

With Pi, he reports completing a product API migration to .NET 10 in about five minutes and shows a successful build. The supplied transcript cuts away before the corresponding oMLX migration result, so it does not establish a precise timing comparison for that task.

The SVG test asks for a pelican riding a bicycle in New York City and compares medium with XHigh reasoning. The speaker reports an MTPLX run of about 1 minute 49 seconds and an oMLX run of about 21 minutes. He prefers the detail in the oMLX image and favors MTPLX for quicker coding work. His conclusion is a tradeoff between responsiveness and perceived output quality in these tests.