Apple's 512GB M5 Ultra can run 93 of 97 open-weight models locally, up to ~50% faster inference
bigaiguy · x · 2026-09-04
The author argues the real story of Apple's M5 Ultra isn't the keynote's "4.3x faster AI" claim but the 512GB unified memory at 1.2TB/s bandwidth in a Mac Studio.
Key numbers:
- M3 Ultra topped out at 819GB/s; since token generation is bandwidth-bound, the jump alone yields roughly 50% faster inference before any GPU gains.
- Per canitrun's tracker estimates (not benchmarks):
- DeepSeek R1 671B @ Q4KM: 12.7 tok/s
- DeepSeek V4 Flash 284B @ Q80: 20.8 tok/s
- Llama 4 Maverick 400B @ Q80: 14.9 tok/s
- GPT-OSS 120B at full BF16: 28 tok/s
- Kimi K2.6 1T @ Q2K: 22 tok/s (fits, but Q2 is a big quality hit)
- 93 of 97 tracked open-weight models fit fully in memory at 8K context. The four that don't are the 1.6T-2.4T giants: DeepSeek V4 Pro, Kimi K3 and Qwen3.8 2.4T.
Bottom line: the 512GB config makes the Mac Studio one of the few consumer devices that can run nearly every major open-weight model locally.
More from Infra
- PostgreSQL 19 slips to November as LLM-assisted code review floods committers with edge cases — sanjaykalra · 2026-10-21
- Decision Models' routing classifier hits AntSeed, cutting inference cost to 24% of always-Opus — airesearch12 · 2026-10-21
- Eon cuts engineer onboarding from 2 days to 2 minutes with an MCP Gateway on Bedrock AgentCore — AWS ML Blog · 2026-10-20
- a16z's Ben Horowitz says AI buildout isn't the 90s bubble — an investor pushes back — henloitsjoyce · 2026-10-20
- Micron and Nvidia drive ~36% of S&P 500's Q3 earnings growth — Beth_Kindig · 2026-10-20
- NVIDIA lands in Europe: Jensen Huang and German chancellor discuss industrial AI at GTC Berlin — nvidia · 2026-10-20