Qwen 3 Flash Next STRATA Q2_0 on M5 Pro Mac Mini 48GB vs M5 Ultra Mac Studio 256 IQ3_S

Summary

The video compares Qwen 3 Flash Next running through Strata on an M5 Pro Mac mini with 48GB of unified memory and an M5 Ultra Mac Studio with 256GB. It is an informative demonstration, but not a strictly controlled comparison: the Mini uses 2-bit quantization, while the Ultra uses IQ3 (roughly 3-bit). The machines also differ substantially in memory capacity and bandwidth.

Performance and memory

• The model is described as roughly 125B parameters and around 70GB in size.
• On the 48GB Mac mini, only part of the model is held in memory; Strata backs the rest with disk storage. The speaker says the Mini can dedicate about 39GB, while the relevant package requires about 46GB.
• Because the package does not fit, Strata disables MTP (multiple-token prediction) on the Mini. The speaker says a 64GB machine should be a better fit for enabling MTP.
• Reported decoding speeds are approximately 22 tokens per second on the Mini and 130 tokens per second on the Ultra. The Ultra also has much higher stated memory bandwidth: about 1.2TB/s, compared with roughly 307GB/s for the Mini.
• Across prompt sizes of roughly 935, 6,771, and 26,000 tokens, the Mini’s measured speed is reported as fairly consistent, at around 19–22 tokens per second.
• The presenter considers the Mini usable, but says they would personally prefer a smaller model on a 48GB machine to leave memory available for other applications and agent tasks. They are interested in testing the model on an M5 Max with 64GB.

Quality and task tests

The presenter emphasizes that outputs can vary in length and detail, so differences in how much each machine produces are not necessarily evidence of a quality difference. Despite its more aggressive quantization, the Mini’s results are described as solid.

Tests shown include:

1. General knowledge: Questions about Tyrannosaurus rex, including bite force, running speed, and skull analysis.
2. Python duration analyzer: The model generated a parser and test cases, then was asked to support fractional durations such as 1.5 or 0.5 days. The presenter checked that the code was run through the interface’s Python interpreter. The Mini eventually showed 28 passing tests.
3. CSV/RAG analysis: Both systems analyzed a 398-row, 9-column automotive dataset, produced summary statistics, and generated reports. The results were reported as accurate.
4. Visualizations: The models created an MPG histogram and a scatter plot. The Mini produced an inline Matplotlib graph; the Ultra also used an ASCII-style plot in one response. The scatter-plot analysis included correlation measures, and the plotted results were described as very similar.
5. LRU cache implementation: Both generated Python implementations, documentation, unit tests, complexity discussions, and a real-world concurrency scenario. The presenter found the implementations broadly comparable and noted that both discussed contention under concurrent access.
6. Follow-up questions: The presenter also checked that the model could continue after earlier coding and knowledge prompts.

Overall takeaway

The Mac mini can run this very large model and handle varied coding, data-analysis, and knowledge tasks, but it is much slower than the Ultra. The speed gap reflects not only the hardware, but also the Mini’s lower memory bandwidth, disk-backed model weights, disabled MTP, and different quantization.

The presenter considered the Mini’s results promising enough to justify further testing, while cautioning that a 48GB system may be better served by a smaller model for everyday multitasking.

Main speaker and sources

• Main speaker: The host of the MLX Compare session from Elite Test Engineering.
• Other sources mentioned: Claude and Codex were used to help prepare the setup documentation; they are not presented as speakers in the video.