Every figure, name and claim this video puts on screen, chased to a primary source.
An MLX native runtime and Mac app for native multi token prediction speculative decoding on Apple Silicon, by youssofal.
The headline claim. The README states: “the model drafts several tokens ahead of itself, verifies each drafted block in a single batched forward pass, and commits tokens through exact rejection sampling with residual correction. Same model, same output distribution, measured 1.6x faster on a 16 GB M4 Mac mini and 2.24x on an M5 Max.” Source: https://github.com/youssofal/MTPLX
M4 Mac mini, the paired figures. The README states: “On a 16 GB M4 Mac mini, tuning the 9B model lands on depth 1: 14.4 tok/s baseline becomes 23.0 tok/s.” 14.4 to 23.0 is 1.597x, which is the 1.6x quoted. Source: https://github.com/youssofal/MTPLX
M5 Max, the fast figure. The Qwen3.6-27B-MTPLX-Optimized-Speed card states: “On the local Apple M5 Max fanmax sustained benchmark, this artifact reached 63.056 tok/s and 62.886 tok/s at depth 3 on the long-code 192-token prompt when using its contract-recommended 3-bit draft-only LM head.” Source: https://huggingface.co/Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed
Its own heads, not a separate draft model. The README describes the MTP head as what enables native speculative decoding: the model drafts its own tokens, with no external draft model required. That is the distinction the script draws against ordinary speculative decoding, which pairs a small drafter with a large verifier. Source: https://github.com/youssofal/MTPLX
Auto tuning. mtplx tune --retune compares plain autoregressive decoding against
depths 1, 2 and 3 on the machine it is run on, and keeps a recommendation only where the
MTP path wins.
Source: https://github.com/youssofal/MTPLX
Which models carry MTP heads. The README states that Qwen 3.5, 3.6 and 3.8 ship with built in MTP heads. Source: https://github.com/youssofal/MTPLX
Apache 2.0, natively multimodal with image and video understanding, tool use and long horizon agentic work, and a native context length of 262,144 tokens. Source: https://huggingface.co/Qwen/Qwen3.8-27B
Up to a 40 core GPU, up to 128 GB unified memory, and up to 614 GB/s unified memory bandwidth. (The script quotes the ceiling of the configurable range, which is what Apple publishes; it is not what every machine ships with.) Sources: https://www.apple.com/newsroom/2026/03/apple-debuts-m5-pro-and-m5-max-to-supercharge-the-most-demanding-pro-workflows/ and https://support.apple.com/en-us/126319
10 core CPU, 10 core GPU, 16 core Neural Engine, 120 GB/s memory bandwidth, 16 GB unified memory as standard. Source: https://support.apple.com/en-us/121555
32 GB GDDR7 on a 512 bit interface. Source: https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
96 GB GDDR7 with ECC, 1792 GB/s memory bandwidth. Source: https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000/
The M5 Max baseline of about 28 tokens per second is not a published figure. The 2.24x multiplier and the 63.056 tok/s depth 3 result are both published; a 28 tok/s starting point is what those two produce when divided, and no source states it directly. The only baseline numbers published on the same benchmark window are greedy diagnostics of the MTPLX artifact itself, between 57.668 and 61.527 tok/s, which measure a different thing. The figure is therefore not drawn on screen. Sources: https://github.com/youssofal/MTPLX and https://huggingface.co/Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed
Acceptance rate is workload dependent. The claim that code accepts better than prose, and that context length, sampler, thermals and model size move the result, is the reasoning the project gives for shipping its tuner. It is not separately measured here.
“Fifty times a day” is a figure of speech, not a measured frequency, and is not put on screen as a number.