i got one....and it's FAST!!!

About this video

The creator benchmarks a loaned Apple M5 Ultra Mac Studio against his own M3 Ultra, treating both as local AI servers with no internet or cloud models. He highlights two spec differences: both have 80 GPU cores but the M5 adds a Neural Accelerator per core, and memory bandwidth rises from 819 gigabytes per second to 1.2 terabytes per second. In his tests, summarizing a 32,000-token prompt with a Qwen 27B 4-bit model, the creator reports the M5 was 3.8 times faster to first token (21.66 seconds versus 82.85 seconds) and 1.5 times faster in generation (43.3 versus 28.1 tokens per second), which he attributes to the accelerators for prompt processing and to memory bandwidth for token generation, since the GPU must reload the model weights for each token. A 14B model widened the margins in his results. He also tests Whisper transcription of a 13.8-minute recording and a 2 hour 5 minute recording across model sizes, video-frame analysis with a Qwen vision model, image generation with FLUX.2, video generation with an LTX model, running a Hermes agent on a local model, a Laya decision-model test classifying 300 texts as spam or not spam, and a local model editing via DaVinci Resolve’s MCP, which he says took about thirteen minutes. He concludes the M5 is strong for daily transcription, image, and video work, notes the configuration he priced at $14,000, and says he wants to cluster future 512-gigabyte models.

Key takeaways

• Creator’s claim: M5 Ultra adds per-core Neural Accelerators and 1.2 TB/s memory bandwidth versus 819 GB/s on M3 Ultra
• Creator’s test: 32K-token Qwen 27B summarization was 3.8x faster to first token and 1.5x faster in tokens per second
• Token generation speed is explained as memory-bandwidth bound because weights are reloaded per token
• Whisper, vision analysis, image and video generation all ran locally and faster on the M5 in his tests
• A local model successfully edited in DaVinci Resolve via MCP, though slowly, in his demonstration

Episode details: published September 23, 2026 · 21:53 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are NetworkChuck’s own claims, shared for education — they are not typical or guaranteed.

← Previous: You Need to Get Your CCNA Right Now!! · Next: you need to try Paperclip RIGHT NOW! →

All NetworkChuck videos · NetworkChuck profile · NetworkChuck on YouTube