# Apple — Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio

- Company: Apple (apple.com)
- Announced: 2026-10-15
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://machinelearning.apple.com/research/mlx-distributed-inference
- Record: https://forck.live/items/22974-local-and-distributed-large-llm-inference-with-mlx-and-rdma-on-mac-studio
- Subject: Machine Learning Research
- Models affected: Muse Spark Open, Mistral Medium 3.5, Qwen 3.8
- Context window: 8,192 tokens

Apple researchers demonstrate running the nearly trillion-parameter Muse Spark Open model on a cluster of four Mac Studios using MLX and RDMA over Thunderbolt, achieving over 36 tokens per second. The M5 Ultra Mac Studio shows up to 4.6x faster time to first token and 1.6x faster generation speed compared to the M3 Ultra, with peak power consumption of about 300W.

## Evidence

Verbatim from https://machinelearning.apple.com/research/mlx-distributed-inference:

> On a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second (see Figure 1 ), well above average human reading speed.

---

Record: https://forck.live/items/22974-local-and-distributed-large-llm-inference-with-mlx-and-rdma-on-mac-studio
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md