MLX-Serve v26.11.2 Runs DeepSeek Across Two Macs, 57% Faster With Tensor Split

HankYeomans · x · 2026-11-07

MLX-Serve v26.11.2 ships multiple ways to run DeepSeek on Macs from 80GB to 512GB of RAM.

• Tensor split across two Macs: the new --dist flag splits DeepSeek-V4.1-Flash over two Thunderbolt 5 Macs working on every token together — two M3 Ultras write 57% faster with DSpark (88 vs 56 tokens/s) and 31% faster without.
• GGUF at MLX speed: with --mlx-gguf, DeepSeek-V4.1 GGUF packs (ISTA GSQ-RCO, Coder, DS4) run on the MLX engine with DSpark drafting and image input, hitting 3,300 tokens/s prompt reading on an M5 Ultra — 2.5x previous. Qwen3, Flint and Muse Glimmer GGUFs also gain faster IQ kernels, and GGUFs now open at their trained context length instead of 32k.
• Other gains: Lossless pack 7% faster on M3 and older, iQ-MLX and expert-pruned packs 25%-43% faster on M4 and older.

Original post →

More from Infra

• JITEngine synthesizes custom LLM inference engines for Apple Silicon, claims up to 2.7x speedup — AccBalanced · 2026-11-07
• Dev predicts 2027's local inference boom: integrated hardware + OS + inference stack, not just GPUs — vincent_koc · 2026-11-07
• vLLM Office Hours Hits 59 Sessions: 76 Speakers from 24 Organizations and Counting — vllm_project · 2026-11-07
• Dev runs 284B-parameter DeepSeek-V4-Flash on CPU only, squeezed into 64GB DDR5 RAM — solyarisoftware · 2026-11-07
• Omnigraph: open-source graph database with git-style branching for multi-agent shared memory — solyarisoftware · 2026-11-07
• Solo dev ships all-Rust bootable OS that runs local LLMs up to 25% faster than llama.cpp — W61k3r · 2026-11-07

Infra channel →