GLM-5.3-Flash on two Mac Studios hits ~140 tokens/s via tensor parallelism
cryotic · reddit · 2026-11-16
Developer cryotic open-sourced GLM-Tandem: running GLM-5.3-Flash tensor-parallel across two Mac Studios, sustaining about 140 tokens/s on fresh prose. The author says the recipe has proven stable after repeated tuning, a useful reference for local LLM deployments.
More from Infra
- Who absorbs failed AI datacenter deals? Microsoft took Abilene's 900MW in three weeks — tengyanAI · 2026-11-16
- Microsoft open-sources TokenEconomizer, cutting coding agent costs 18–31% — aigclink · 2026-11-16
- EE asks Claude to price 1,000 human brains in silicon: 39 sq-mile data center, 27x global nuclear output — Recent-Day3062 · 2026-11-16
- Intel 170hx cards unlocked to 80GB VRAM, price jumps from hundreds to $4k — GabGarrett · 2026-11-16
- Track cost per finished task, not token price: the agent metric that matters — arpit_bhayani · 2026-11-16
- Wuray's WQS: 24 GPUs + one storage node match 48-GPU token output, raises $14M+ — 创业邦 · 2026-11-16