---
type: "post"
id: "mlx-serve-v26-11-2-adds-two-mac-3255a5"
slug: "mlx-serve-v26-11-2-adds-two-mac-3255a5"
title: "mlx-serve v26.11.2 adds two-Mac DeepSeek inference and faster GGUF support"
description: "mlx-serve v26.11.2 adds distributed inference across two Thunderbolt 5 Macs and an MLX-backed path for DeepSeek-V4.1 GGUF models."
retrieval_nugget: "mlx-serve v26.11.2 adds distributed inference across two Thunderbolt 5 Macs and an MLX-backed path for DeepSeek-V4.1 GGUF models. The release reports faster generation and prompt processing on specific Apple hardware, alongside caching, memory, and compatibility changes. For teams evaluating local inference, it offers new options to test—not universal performance guarantees."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["inference"]
entities: ["mlx-serve","DeepSeek-V4.1-Flash","DeepSeek-V4.1","MLX","Apple","Qwen3"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "mlx-serve v26.11.2 adds distributed DeepSeek inference across two Thunderbolt 5 Macs and faster GGUF support."
prior_refs: ["https://newruntime.com/posts/mlx-serve-v26-11-1-adds-installable-model-3f3b35/"]
source_url: "https://github.com/ddalcu/mlx-serve/releases/tag/v26.11.2"
source_urls: ["https://github.com/ddalcu/mlx-serve/releases/tag/v26.11.2"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:mlx-serve-v26-11-2-adds-two-mac-3255a5"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/mlx-serve-v26-11-2-adds-two-mac-3255a5/","markdown":"https://newruntime.com/posts/mlx-serve-v26-11-2-adds-two-mac-3255a5.md","json":"https://newruntime.com/posts/mlx-serve-v26-11-2-adds-two-mac-3255a5.json"}
markdown_url: "https://newruntime.com/posts/mlx-serve-v26-11-2-adds-two-mac-3255a5.md"
---

# mlx-serve v26.11.2 adds two-Mac DeepSeek inference and faster GGUF support

## Retrieval answer

mlx-serve v26.11.2 adds distributed inference across two Thunderbolt 5 Macs and an MLX-backed path for DeepSeek-V4.1 GGUF models. The release reports faster generation and prompt processing on specific Apple hardware, alongside caching, memory, and compatibility changes. For teams evaluating local inference, it offers new options to test—not universal performance guarantees.

## Summary
mlx-serve v26.11.2 adds distributed inference across two Thunderbolt 5 Macs and an MLX-backed path for DeepSeek-V4.1 GGUF models. The release reports faster generation and prompt processing on specific Apple hardware, alongside caching, memory, and compatibility changes. For teams evaluating local inference, it offers new options to test—not universal performance guarantees.

## What happened
The mlx-serve project released v26.11.2 on November 6. Its headline addition, `--dist`, splits a DeepSeek-V4.1-Flash model across two Macs connected over Thunderbolt 5, with both machines contributing to each generated token. On two M3 Ultras, the project reports generation 57% faster with DSpark drafting enabled: 88 tokens per second rather than 56. Without DSpark, it reports a 31% speed increase. The release does not give the exact non-DSpark tokens-per-second figures.

A second major change, `--mlx-gguf`, runs specified DeepSeek-V4.1 GGUF model sets on the MLX engine. The release names ISTA’s GSQ-RCO packs, a Coder pack, and ivanfioravanti’s DS4 packs. This path supports DSpark drafting and image input; the project says a llama.cpp image projector placed beside the model files can also be used. On an M5 Ultra, it reports about 3,300 tokens per second for reading prompts, 2.5 times its previous speed and close to its own MLX pack. The release also says the faster GGUF route covers Qwen3, Flint, and Muse Glimmer, with faster IQ kernels. Separately, GGUFs using the built-in DeepSeek engine now open at their trained context length rather than being limited to 32,000 tokens.

The release describes further performance and state-management changes. DeepSeek-V4.1’s Lossless pack writes 7% faster on M3 and older Macs; its iQ-MLX and expert-pruned packs write 25% to 43% faster on M4 and older Macs. Three-bit mixture-of-experts models such as Qwen3-Coder-30B-A3B write a little faster. Models stay in memory between requests rather than being squeezed out by macOS, which the project says could otherwise freeze a Mac running a model near its memory size.

For models that retain their own state, including DeepSeek-V4.1, conversations now keep a few earlier checkpoints. Editing or forking an older message at 60,000 tokens reportedly takes 0.13 seconds instead of 8.7 seconds. A long Qwen3.8-Flash-Next conversation moved to an SSD reportedly resumes from a checkpoint in five seconds instead of 46. The release also updates mlx-serve to MLX 0.32.5 and lists plugin, app, model-switching, and compatibility fixes. It does not state a price or describe a hosted service; these are local runtime capabilities and changes.

## What the archive adds
New Runtime's earlier coverage examined MacSeek’s local DeepSeek-V4.1 GGUF inference and oMLX’s Apple-silicon optimizations, both dated 2026-10-31: [MacSeek](https://newruntime.com/posts/macseek-from-scratch-metal-engine-runs-the-233-005b3d/) and [oMLX](https://newruntime.com/posts/omlx-0-7-3-speeds-up-deepseek-v4-fd0577/). This is a new event in the same local-inference thread, with new runtime capabilities and performance claims rather than a restatement of those projects.

## Trend signal
This adds another example of teams adapting local inference to Apple hardware through model formats, engine choices, memory handling, and multi-machine execution. The release’s figures are project-reported results for particular hardware and configurations, not a general benchmark. This does not yet fit a pattern New Runtime tracks.

## Connections to other events
The release meets New Runtime's earlier coverage of MacSeek, which reported running a 233 GB DeepSeek-V4.1-Flash GGUF on a 64 GB Mac, and oMLX, which described DeepSeek V4.1 prefill and concurrent-request changes on Apple silicon. Those notes concern different projects and approaches; mlx-serve adds a two-Mac distribution mode and an MLX execution path for named GGUF packs.

## Why it matters
Local inference is often constrained by memory capacity, model format support, and the time required to process long prompts or continue a conversation. This release targets each of those practical bottlenecks: splitting a model across two Macs may make a larger workload feasible on connected machines, while the GGUF path could let teams use existing model packs with MLX execution and image input. Checkpoint reuse could also make revisiting an old point in a long conversation less costly in time.

The reported speedups may be useful when deciding what to test, but they do not establish that a two-Mac setup is cheaper, more reliable, or faster for every workload. Teams should compare the exact model pack, hardware, context length, and drafting configuration they plan to run. The update is most relevant to developers already experimenting with mlx-serve and Apple-silicon inference, not to users seeking a general hosted model API.

## Limits and uncertainty
The performance figures are claims in the project’s release notes and have not been independently verified here. The release gives selected hardware and some configurations, but does not provide a complete benchmark methodology, repeated-run results, power use, or end-to-end latency distributions. It also does not quantify the cost or operational trade-offs of connecting two Macs, or specify the limits of distributed execution across other hardware combinations. Compatibility and image-input behavior may vary by model pack and setup; the release does not establish that every GGUF model works through the new path.

## Skill to build now
If you already run mlx-serve, test the new paths against one representative workload before changing a production setup. Record the exact model pack, Mac hardware, context length, whether DSpark is enabled, prompt-processing speed, generation speed, and memory behavior. For long chats, compare the time to fork or edit an earlier turn at a fixed conversation length. This makes the release’s specific claims actionable without treating them as a promise for your own workload.

## Reader takeaway
mlx-serve v26.11.2 makes local DeepSeek inference more flexible for some Apple-silicon setups: it can distribute DeepSeek-V4.1-Flash across two Macs and run named GGUF packs through MLX. Its speed figures are configuration-specific project claims. Treat the update as a reason to test against your actual model, hardware, and conversation pattern.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
