Apple AI Updates: October 16, 2026

1. New MLX Release Runs Meta’s 816B Muse Spark Open Across Networked Mac Studios

Apple. Apple released a new version of MLX alongside a post showing Meta’s Muse Spark Open (816B total, 42B active parameters) running locally on Mac Studio clusters connected by RDMA over Thunderbolt 5. In NVFP4 the model needs about 493GB for weights alone, so it is split across machines using pipeline or tensor parallelization through MLX’s jaccl backend, reaching over 36 tokens per second on four M5 Ultra Mac Studios at an 8,192-token context. On single-machine benchmarks with Qwen 3.8 27B and Mistral Medium 3.5 128B, the M5 Ultra delivered up to 4.6x faster time to first token and 1.6x faster decoding than the M3 Ultra, at a peak power draw of about 300W. Source

2. Selection-Based Structured Reasoning Cuts Per-Turn Reasoning Latency for Small Search Agents by Over 90%

Apple. Apple researchers propose Selection-based Structured Reasoning (SSR), which replaces free-form reasoning before each agent action with selection from a set of pre-written, reusable reasoning candidates scored by likelihood. Because candidates are fixed, they can be scored in parallel with a shared KV cache. Across seven multimodal search benchmarks with 2B and 4B models, SSR matched leading same-size search agents on success rate while cutting per-turn reasoning latency by more than 90% and total per-question inference latency by 28% to 54%. Source

3. EMNLP Study Finds Plan-Then-Retrieve Tool Discovery Wins Only on Multi-Tool Queries

Apple. An EMNLP paper compares the common Retrieve-then-Plan approach for finding tools in large API registries with Plan-then-Retrieve, where the LLM first writes hypothetical function stubs used as retrieval queries. On ToolBench (46,984 APIs), Retrieve-then-Plan performed better on single-tool queries and Plan-then-Retrieve performed better on multi-tool queries, a pattern that held across retrieval budgets, registry sizes from 1k to 47k APIs, and both Gemini 2.5 Flash and gpt-oss-120b. Plan-then-Retrieve roughly doubles inference cost, and the authors suggest routing by query complexity using the top-1 retrieval score as a signal. Source