YOU ARE VIEWING ONE ITEM FROM THE AICRIER FEED

LM Studio M5 speeds look uneven

AICrier tracks AI developer news across Product Hunt, GitHub, Hacker News, YouTube, X, arXiv, and more. This page keeps the article you opened front and center while giving you a path into the live feed.

// WHAT AICRIER DOES

7+

TRACKED FEEDS

24/7

SCRAPED FEED

Short summaries, external links, screenshots, relevance scoring, tags, and featured picks for AI builders.

LM Studio M5 speeds look uneven
OPEN LINK ↗
// 59d agoBENCHMARK RESULT

LM Studio M5 speeds look uneven

A user with a 32GB M5 MacBook Pro is sanity-checking LM Studio throughput after seeing 8 t/s on Gemma 3 27B 4-bit MLX, 32 t/s on Nemotron 3 Nano 4B GGUF, and 39 t/s on GPT OSS 20B MLX under default context settings. The thread is a call for comparable M5 or MacBook Air/Pro numbers to see whether the slowdown is normal or a tuning issue.

// ANALYSIS

Local AI speed on Apple Silicon is a moving target, and this reads less like a broken machine than a calibration check for LM Studio's newest runtime path.

  • LM Studio officially supports both `llama.cpp` GGUF and Apple `MLX` on Apple Silicon, and Apple calls out LM Studio as one of the apps that should benefit from M5's Neural Accelerators.
  • LM Studio has already shipped M5-specific MLX NAX auto-upgrade fixes, so benchmark comparisons on this chip age fast and older runtime builds can undershoot.
  • Community M5 reports already show Nemotron nano 4bit around ~55 t/s after a runtime switch, which makes the poster's 32 t/s believable but not especially fast.
  • The 27B Gemma run is the outlier: 8 t/s points to a bandwidth-heavy workload, a heavier context, or a model/backend pairing that is not hitting the chip's best path.
// TAGS
lm-studiollminferencebenchmarkgpuself-hosted

DISCOVERED

59d ago

2026-03-29

PUBLISHED

59d ago

2026-03-29

RELEVANCE

8/ 10

AUTHOR

nemuro87