Dip your toe into local AI at Micro Center, the AI destination:
Austin, Texas: sign up now for a FREE 128GB flash drive, redeemable when the new Austin store opens:
Apple says the new M5 Ultra Mac Studio is up to 4x faster than the M3 Ultra at LLM prompt processing. I own an M3 Ultra Mac Studio, so I raced them head to head on local AI: same model, same prompt, same settings. On a 32,000-token prompt with Qwen3.8 27B, the M5 Ultra started answering in 21.7 seconds. The M3 Ultra took 82.8. That’s 3.8x. Then the M5 wrote the answer 1.5x faster (43.3 vs 28.1 tokens per second), which lines up with its jump in memory bandwidth from 819 GB/s to 1.2 TB/s.
Then I threw my real work at it. Whisper transcribed two hours of my audio in about a minute and a half. A local vision model described my footage, a job I pay the Gemini video API for right now. I generated images with FLUX.2 and video with LTX-2.5, ran my Hermes agent on a local model, tried an open decision model (Laya) in place of Jev, and let a local model edit a video in DaVinci Resolve through its MCP server. All of it runs on the Mac. No internet required.
In this video you’ll learn how the M5 Ultra and M3 Ultra compare for running LLMs locally with MLX, why the wait for the first word (time to first token) got about 4x shorter while writing speed only got 1.5x faster, how model size and dense vs mixture-of-experts models change the numbers, and whether a Mac Studio makes sense as your own local AI server.
Apple loaned me the M5 Ultra Mac Studio for this video, and it goes back to them. They don’t control what I say, and all the tests and all the numbers are mine.
Join the NetworkChuck Academy!:
RESOURCES / LINKS:
📊 Every result, the raw data and my benchmark tool:
🌐 The results as a web page:
🍎 Shop Apple at Micro Center:
📰 Micro Center News:
🛠️ MLX (Apple’s machine learning framework):
🛠️ mlx-lm (runs the chat models):
🤖 Qwen3.8 27B, 4-bit MLX (the race model):
🎙️ Whisper on MLX:
🖼️ mflux (FLUX.2 on MLX):
🖼️ FLUX.2 klein 4B:
🎬 LTX-2.5 video model:
🧠 Laya decision model:
🤖 Hermes Agent:
📺 Set up your own Hermes agent:
🤖 LM Studio (and Bionic):
🎬 DaVinci Resolve:
**Sponsored by Micro Center
SUPPORT NETWORKCHUCK:
☕☕ COFFEE and MERCH:
READY TO LEARN??
🔥🔥Join the NetworkChuck Academy!:
📚 CCNA Course:
FOLLOW ME EVERYWHERE:
Instagram:
X/Twitter:
Facebook:
Join the Discord server:
Some links in this description are affiliate links. If you buy through them, I may earn a small commission at no extra cost to you.
#macstudio #localai #m5ultra
source
Dip your toe into local AI at Micro Center, the AI destination: https://micro.center/322535
Austin, Texas: sign up now for a FREE 128GB flash drive, redeemable when the new Austin store opens: https://micro.center/196cca
Apple says the new M5 Ultra Mac Studio is up to 4x faster than the M3 Ultra at LLM prompt processing. I own an M3 Ultra Mac Studio, so I raced them head to head on local AI: same model, same prompt, same settings. On a 32,000-token prompt with Qwen3.8 27B, the M5 Ultra started answering in 21.7 seconds. The M3 Ultra took 82.8. That's 3.8x. Then the M5 wrote the answer 1.5x faster (43.3 vs 28.1 tokens per second), which lines up with its jump in memory bandwidth from 819 GB/s to 1.2 TB/s.
Then I threw my real work at it. Whisper transcribed two hours of my audio in about a minute and a half. A local vision model described my footage, a job I pay the Gemini video API for right now. I generated images with FLUX.2 and video with LTX-2.5, ran my Hermes agent on a local model, tried an open decision model (Laya) in place of Jev, and let a local model edit a video in DaVinci Resolve through its MCP server. All of it runs on the Mac. No internet required.
In this video you'll learn how the M5 Ultra and M3 Ultra compare for running LLMs locally with MLX, why the wait for the first word (time to first token) got about 4x shorter while writing speed only got 1.5x faster, how model size and dense vs mixture-of-experts models change the numbers, and whether a Mac Studio makes sense as your own local AI server.
Apple loaned me the M5 Ultra Mac Studio for this video, and it goes back to them. They don't control what I say, and all the tests and all the numbers are mine.
Join the NetworkChuck Academy!: https://ntck.co/NCAcademy
dude why arent u running qwen3.8-flash-next:125b-mlx arghhhhhh!!! i'm just at the beginning of the video…tell me u didnt skip qwen3.8-flash-next:125b-mlx
3 is too expensive for local AI. You can pay a Claude max subscription for more than 10 years for the same price.
Thank you for your content and prayer.
Is there enough of a speed difference with the 96GB version of the M5 Ultra compared to the 256GB version? Would speeds be similar the the 256GB vs the M3 Ultra? That would be a cool test…. The 256 is just really priced out of my league.
run claude with ollama
go try glm falsh 5.3
Loved the end Chuck! Thank you and keep it up!
I was surprised and so thankful for you to give us all such a heartfelt prayer! Came for the info and left with a blessing – thank you!
Thanks for the prays man love it and appreciate the pray and you info. Be blessed
My M3U uses an 8-bit Qwen 3.8 27B and averages 38, topping out at 55 on MTPLX…my M5U arrives next week
I’ve been using local models on my m4pro Mac mini for my agents. Worked well. I switched to my M5max Mac Studio. Works perfectly. Running Ornith1.5:35b
I absolutely love that you started praying at the end of your videos. I've never seen anyone else do that, and I'm a big fan. Keep on going!
what is the minimum requirements specifications for running local llm?
would love one of those, but as a student i cant even dream about the money for one x)
Thank you, and looking forward to get one. God bless you too.
Im new to ai and i keep looking for videos on youtube to learn. Thats how i stumbled upon this video and let me tell you, i really needed that prayer more than you can imagine. Thank you brother. You have a new subscriber
Your own personal Jesus running on m5 ultra
I really appreciated your prayers. It is genuine and sincere and human which we need more of these days
I am currious to see DGX vs M5 ultra for agents, multiple agents and especially long context type stuff.
Learning tech and God as a Chirstian bro in tech. So amazing! That prayer got me. First time seeing it. I can't stop smiling.- like so glad. Love you!✝
The prayer earned a subscriber and I will definitely be recommending you to others because of it.
looking forward to the m5 max cluster… can't wait to watch that video, very interested in seeing those bench marks and how it works
21 seconds for TTFT is kinda a joke for a model of this size.
I can send 100k tokens at Qwen 27B with a 3090 and get a decoded response back in less than 10 seconds. These little AI boxes are a waste of money..
Nothing weird about that, thank you so much!
Do some cable management pleaseeee Chuck!
Awesome vid annyway 🙂
For $14,000 I'd put that towards a house instead unless that $14,000 is making me money!
God bless you and God shine his light upon you…
Keep making these!! Love this content.
Chuck, God bless you!
That would be a fairly used car so ima say I cant afford that toy lol
Thanks for the great video!
Godbless You Brother Love The Videos ‼️💕
better to had 2 DGX spark with the benchmark 😉
4 Mac Studio M5 Ultra, 512 GB, all connected – logical next step. Do it.
Amen. You are different and I like that. I always wanted to have my own personal local AI. But the hardware is crazy expensive. I have watched several times your videos and I I have used some of your skills. Continue the hard work bro. Also praying for you
I know this is crazy, but can I get one of the 1000 old Mac Studios you don’t use so I can run some projects I’ve been wanting to try 🥹🥹🥹
"And it's free" ngl $14k must be the most expensive free I've heard recently ahah