whatcani.run의 실사용 데이터(22,914,944 토큰·4,479회·191명)를 바탕으로 M1 Max(64GB)에서 로컬로 돌릴 수 있는 모델 성능을 정리했습니다. llama.cpp·mlx_lm 등으로 측정한 결과, 1B–4B급 모델은 메모리 0.6–4.6GB로 'runs great/well', 4–13GB대 모델은 'runs well/ok', 20–26B급(예: gpt-oss-20b, Gemma 26B)은 11–13GB로 간헐적 실행. Qwen 계열과 Liquid AI 모델이 소형 환경에서 특히 우수했습니다.

https://www.whatcani.run/

#macos #m1max #localllms #benchmarks #qwen

whatcani.run

Find the best models and how to run them locally.

whatcani.run

Alex Cheema (@alexocheema)

4년 된 M1 Max Mac Studio 6대를 재활용해 MiniMax AI의 M2.5를 Exo Labs로 실행하는 사례가 공유됐다. 중고 장비와 높은 메모리 대역폭을 활용해 저비용으로 대규모 AI 클러스터를 구성할 수 있음을 보여준다.

https://x.com/alexocheema/status/2037001428758135192

#minimax #exolabs #macstudio #aicluster #m1max

Alex Cheema (@alexocheema) on X

6 x M1 Max mac studios repurposed to run @MiniMax_AI M2.5 using @exolabs. These are 4 year old devices, each with 400GB/s memory bandwidth (total 2.4TB/s). Second hand each mac is ~$1.2k.

X (formerly Twitter)

Awni Hannun (@awnihannun)

M5 Max가 노트북 폼팩터에서 로컬 AI 처리 성능을 크게 끌어올린 신제품이라는 내용입니다. M1 Max 대비 최대 8배 빠른 prefill 및 이미지 생성 속도를 보였고, 벤치마크는 MLX / mlx-lm으로 수행되었습니다. 로컬 LLM·이미지 생성 워크로드에서 성능 향상을 기대할 수 있습니다.

https://x.com/awnihannun/status/2028852360190345687

#m5max #m1max #mlx #mlxlm

Awni Hannun (@awnihannun) on X

M5 Max is a local AI powerhouse in a laptop form factor. So awesome to see this thing released. Up to 8x faster prefill / image generation compared to M1 Max. Benchmarks done with MLX / mlx-lm.

X (formerly Twitter)
@Transcedental @9to5Mac yes and they used to care about quality. Those days are gone. I read that out of 20 machines that upgraded to #macos26 6 were bricked in a single building. #macos26.0.1 also bricked devices. The first bricked my #m1max. Thankfully I have the #M4Max. The fact #apple is doing ANYTHING other than focusing on this is shameful. They even knew about it in beta testing. There is no excuse for this.
@madeindex the #MacBookPro #M4Max has an external (USB-C) HDD for backup. The #M1Max is bricked thanks to #Apple releasing their crapware #macOS (#macOS26) knowing full well that in beta they were bricking devices. But once I get that repaired (DFU failed here) if Apple's so-called 'geniuses' can fix it it's still not worth an extra backup as it was used to set up the M4 Max. The M1 Max needs the keyboard replaced too. Thank goodness I got the M4 Max earlier this year!

Using #Atari8Bit based #SynapseSoftware #SynCalc to compare #GeekBench values and decide on Apple #MacBookPro #M1Max 10/32 upgrade path. M4 Max seems to be the only viable path.

The A8 posed a bit of struggle with limited screen space.

#retrocomputing #vintagecomputing

Using #AtariST based #LDWPower to compare #GeekBench values and decide on Apple #MacBookPro #M1Max 10/32 upgrade path. M4 Max seems to be the only viable path.
#retrocomputing #vintagecomputing
Using #MSDOS based #QuattroPro 5 to compare #GeekBench values and decide on Apple #MacBookPro #M1Max 10/32 upgrade path. M4 Max seems to be the only viable path.

@BasicAppleGuy stickers have arrived!

#apple #mac #m4pro #m1max

@paul Looking forward to my #M4PRO in a few days and comparison with the #M1Max. Curious how real world performance will compare.