Citation and evidence

exo (software)

26 min full readUpdated 51 references

This article's verification

Report a problem with this article

Edit article
More

Cite this article

exo (styled in lower case; the app and release notes write EXO) is open-source software that joins several of a user's own computers into one cluster for running AI models. It is developed by the British company EXO Labs, and its README opens with the sentence "exo connects all your devices into an AI cluster."[1][2] The stated purpose is twofold: to run large language models that do not fit in one machine's memory by splitting them across devices, and, on Macs cabled together with Thunderbolt 5, to make a model run faster as devices are added.[1] The models it runs are open-weight models downloaded from Hugging Face.[1]

The project has had two distinct designs. The first, published in 2024 under the GNU General Public License version 3, was a peer-to-peer Python program that ran on Macs, Linux machines and single-board computers, using MLX and tinygrad as inference engines.[3][4] The second, called EXO 1.0, was merged into the public repository on 18 December 2025, and an Apache License 2.0 file was added to the repository later the same day. It is a rewrite that uses MLX only, elects a coordinating node, and lists Apple silicon Macs as its only tested platform.[5][6][7][8] Both licenses are approved by the Open Source Initiative.[9][10]

The code is hosted in the GitHub repository exo-explore/exo, which was created on 2024-06-24; the first version's README gave the same address.[11][3] On 19 November 2026 the repository had 48,492 stars and 3,630 forks.[11] The most recent release at that date was v1.0.71, published on 2026-04-23, and the most recent commit on the main branch was dated 2026-11-06.[12][13]

History

2024: the first version

The earliest commits in the repository are dated 23 and 24 June 2024 and were written by Alex Cheema, a co-founder of EXO Labs.[14][15] A GPL version 3 license file was added on 2024-07-16.[4] VentureBeat reported in November 2024 that the company had released the software on GitHub in July of that year.[15]

The README of that version, cited here as it stood at its last revision on 2025-10-30, carried the tagline "Run your own AI cluster at home with everyday devices" and invited users to combine an "iPhone, iPad, Android, Mac, NVIDIA, Raspberry Pi, pretty much any device".[3] It stated that "exo does not use a master-worker architecture" and that devices connected to each other as peers. The default way of dividing a model, "ring memory weighted partitioning", ran inference around a ring in which each device ran a number of model layers proportional to its memory. MLX was the inference engine on Macs and tinygrad on Linux, and the README said the two were interoperable within one cluster; PyTorch and llama.cpp engines were listed as unfinished. Devices found each other by UDP, a manual list or Tailscale, and exchanged data over gRPC. The program offered a ChatGPT-compatible API and a web chat page based on tinygrad's tinychat, both on port 52415.[3]

The stated hardware requirement was only that the devices together hold enough memory for the whole model. The README also labeled the software experimental, and it said the iOS implementation had fallen behind the Python code and was not being distributed.[3]

In November 2024 VentureBeat described a cluster that Cheema had built from four Mac Mini M4 computers and one MacBook Pro M4 Max, with a retail cost it put at about $5,000, and reported the token rates given in the benchmarks table below.[15] From 25 December 2024 to 5 January 2025 the company published a daily blog series, "12 Days of EXO", that included benchmark results, a run of DeepSeek V3 on eight Mac Minis, and posts on distributed training.[16][17] The post of 31 December 2024 said: "Early in 2025, EXO will be moving to a more open license."[18]

2025: a pause in public work and a rewrite

The last tag of the first version, v0.0.15-alpha, points to a commit dated 2025-02-25.[19] On 14 April 2025 a user opened an issue titled "Exo Project Status", writing that there had been almost no commits or replies for some time. The issue stayed open until December, and commenters in it pointed to community forks.[20] On 21 October 2025 a notice was added to the README that read in part: "What began here is becoming part of something much larger."[21]

During this period EXO Labs showed the rewritten software before releasing it. A company blog post of 15 October 2025 described "EXO 1.0" running one request across an NVIDIA DGX Spark and a Mac Studio.[22] Tom's Hardware, covering that post the next day, noted that the public version was still 0.0.15-alpha and that the 1.0 build "isn't public yet".[23] In November 2025 Engadget reported watching a demonstration in which four Mac Studios ran the Kimi K2 Thinking model "in an early version" of EXO 1.0, using a Thunderbolt 5 clustering feature that Apple was adding in macOS Tahoe 26.2. The article put the cluster's power draw at under 500 watts.[24] heise online reported the same demonstration.[2]

The rewrite reached the public repository in December 2025. A commit of 17 December titled "prep repo for v1" removed every file of the first version except the README, including the GPL license file, and a merge commit titled "exo v1" brought in the rewritten code on 18 December.[25][5] The merged files did not include a license file. A commit titled "re-add LICENSE" added the Apache License 2.0 text later that day, and a further commit the same day replaced the template's placeholder copyright line with "Copyright 2025 Exo Technologies Ltd".[5][6][45] The earlier code was kept in a separate, archived repository, exo-explore/ex-exo, whose license file is the GPL version 3.[26] The status issue was closed on 18 December by the GitHub account Evanev7, which GitHub labels a collaborator on the repository, with the comment "EXO 1.0 is now released under Apache 2.0!"[20]

2026: releases of the 1.0 line

Packaged releases of the new version began with alpha builds on 22 December 2025. The first non-alpha release, v1.0.61, was published on 2026-01-08.[12]

ReleaseDateChanges named in the release notes
v1.0.612026-01-08Packages the original v1 release with small fixes; exo-bench benchmarking tool; option to run without a worker [27]
v1.0.632026-01-16Stability fixes for long-running clusters; uninstall button in the macOS app [28]
v1.0.682026-02-25Custom models from Hugging Face; Claude Messages, OpenAI Responses and Ollama APIs; experimental image generation [29]
v1.0.692026-03-27Continuous batching on by default; support for M5 Pro and M5 Max chips [30]
v1.0.702026-04-17Multimodality; lower memory use with long contexts [31]
v1.0.712026-04-23Sampling defaults; fixes for M5 Macs and RDMA; support for Kimi K2.6 [32]

Expanded article table

No release was published between 23 April and 19 November 2026.[12] On the main branch, a commit of 2026-06-03 replaced the libp2p networking layer with zenoh.[33] The branch's history then shows commits on 22 June, a single commit on 25 August, and no others until 12 October 2026, when commits resumed.[50] The latest commit at the time of writing, dated 2026-11-06, changed request handling so that, in the words of its title, each request "streams over HTTP from the device that runs it".[13] The v1.0.71 release predates all of these changes.[12][33] The README on the main branch, last changed in November 2026, is newer than the latest release. It notes that the packaged app uses its release number (it gives 1.0.71 as an example) while a source checkout uses the version in pyproject.toml (0.3.70).[1][34]

The two designs compared

AspectFirst version (2024 to 2025)EXO 1.0 line (from December 2025)
LicenseGNU GPL version 3 [4]Apache License 2.0 [6]
CoordinationPeers, no master [3]Elected master, event-sourced state [7]
Inference engineMLX and tinygrad [3]MLX [1]
NetworkingUDP, manual or Tailscale discovery; gRPC [3]Automatic discovery; zenoh on the main branch since June 2026, libp2p before [1][33]
Model splitRing weighted by device memory [3]Pipeline or tensor sharding, chosen from the topology [1]
API formatsChatGPT-compatible [3]OpenAI Chat Completions and Responses, Claude Messages, Ollama [1]
Hardware named in the documentationiPhone, iPad, Android, Mac, NVIDIA, Raspberry Pi [3]Apple silicon Macs as tested hardware; Linux from source on CPU [8][1]

Expanded article table

How the current version works

This section follows the documentation on the main branch as of November 2026, which is ahead of the latest packaged release (v1.0.71, April 2026).[1][12]

Nodes and coordination

Every device runs the same program, called a node in the repository's AGENTS.md guide. A node contains a router for publish-subscribe messaging (zenoh, reached through Rust bindings), a worker that downloads models and runs inference, a master that keeps the cluster's state and places model instances on nodes, an election component that chooses the master with the bully algorithm, and an HTTP API built on FastAPI.[7] State is kept by event sourcing: the master orders events and broadcasts them, and workers apply them. A node that has fallen too far behind receives a snapshot of the state.[7] This is a change from the first version, which had no master.[3] The commit of 6 November 2026 narrowed the master's part in serving requests: its message says that no request enters the cluster's log and that "The master only places instances".[13]

Devices on the same network find each other without configuration, but only if they run the same version of exo. The README lists further conditions: the same cluster name, the same "wire version" and the same build. As a result a Mac and a Linux machine do not join each other, and neither do the Mac app and a source checkout of the same release.[1] Each node serves the API and a web dashboard at http://localhost:52415.[1] The dashboard is written in Svelte and TypeScript, and the Mac app runs in the menu bar.[7][1]

Splitting a model

The README says exo chooses how to split a model from "a realtime view of your device topology", taking into account each device's resources and the latency and bandwidth of each link.[1] A client can ask for the possible placements of a model before launching it through the /instance/previews endpoint.[1][35] The benchmarking tool's options show the two choices a placement combines: the sharding method, pipeline or tensor, and the communication backend, ring or jaccl.[1]

EXO Labs explained pipeline sharding in a December 2024 post about the first version. The model's layers are cut into contiguous slices called shards, one per device. A request passes through the shards in order, and each device sends its output activations to the next. The post said these activations are small (under 4 KB for Llama 3.2 3B), so that latency between devices, and not bandwidth, is usually the limit. It called the technique "pipeline parallel inference", an inference-time use of pipeline parallelism.[36]

Tensor parallelism is the method the README associates with speed gains. It claims "up to 1.8x speedup on 2 devices and 3.2x speedup on 4 devices".[1] Release notes record tensor sharding being added for individual model families, for example gpt-oss in v1.0.63.[28]

For communication between devices exo uses MLX's distributed layer.[1] MLX documents a ring backend that works over TCP sockets and a backend called JACCL that uses RDMA over Thunderbolt, which its documentation calls necessary for tensor parallelism.[37]

RDMA over Thunderbolt 5

Remote direct memory access over Thunderbolt is a feature of macOS Tahoe 26.2. Apple's release notes describe it as enabling "low-latency communication between Thunderbolt 5 hosts for use cases including distributed AI inference using MLX".[38] MLX's documentation says the JACCL backend built on it reaches a communication latency "an order of magnitude lower than the ring backend".[37] The exo README claims a "99% reduction in latency between devices".[1] Jeff Geerling, who tested a four-Mac cluster on loan from Apple, wrote that RDMA lowered memory access latency from 300 microseconds to under 50.[39]

RDMA has to be enabled by hand on each Mac, and MLX's documentation says this cannot be done remotely.[37] The README's procedure is to boot the Mac into Recovery mode, open Terminal and run:

rdma_ctl enable

It lists the Macs with Thunderbolt 5 as the M4 Pro Mac Mini, the M4 Max Mac Studio, the M4 Max MacBook Pro and the M3 Ultra Mac Studio.[1]

APIs and models

Since v1.0.68 a node answers in four request formats: the OpenAI Chat Completions format at /v1/chat/completions, Anthropic's Claude Messages format at /v1/messages, the OpenAI Responses API format at /v1/responses, and the Ollama API under /ollama/api/.[1][29] As with other servers that offer an OpenAI-compatible API, the API documentation says existing clients can be pointed at exo by changing the base URL.[35] The release notes name Claude Code as a tool the Messages endpoint enables, and the README names Open WebUI for the Ollama endpoint.[29][1]

Models are fetched from Hugging Face; the README's examples use MLX-format repositories such as mlx-community/Llama-3.2-1B-Instruct-4bit.[1] Other Hugging Face models can be registered through /models/add. A model whose configuration requires trust_remote_code is not loaded unless the user enables it explicitly, because such a model runs code from its repository.[1][35] The API reference documents image generation and editing endpoints. The README lists image model support as off by default, turned on with the EXO_ENABLE_IMAGE_MODELS environment variable, and the v1.0.68 notes call image generation experimental.[35][1][29]

Installation and requirements

The Mac app is distributed as a disk image from EXO Labs and as a Homebrew cask, installed with brew install --cask exo.[1][40] The README gives these commands for running from source on macOS:

# Clone exo
git clone https://github.com/exo-explore/exo

# Build dashboard
cd exo/dashboard && npm install && npm run build && cd ..

# Install Python dependencies, including the MLX backend
uv sync --extra mlx

# Run exo
uv run exo

On Linux the dependency step is uv sync --extra mlx-cpu, with mlx-cuda13 and mlx-cuda12 extras noted for NVIDIA hardware.[1]

ItemDocumented requirement or status, November 2026
Mac appApple silicon Mac with macOS Sequoia 15.1 or later, according to the README; the company website gives a different minimum (see below) [1][41]
RDMA over ThunderboltmacOS Tahoe 26.2 or later, Thunderbolt 5 ports and cables [1]
Source build on macOSXcode, Homebrew, uv, Node and Rust; Python 3.13 [1][34]
Source build on Linuxuv, Node 18 or higher and Rust [1]
AccelerationGPU on macOS; CPU on Linux, with GPU support "under development" [1]
Tested hardware ("Tier 1")Mac Studio M3 Ultra, Mac Mini M4 Pro, MacBook Pro M5 and M4 Max [8]
Planned platforms (list dated 2025-12-18)Linux with CUDA (DGX Spark) and Linux CPU; Windows under the heading "Longer term!" [8]

Expanded article table

The first-party sources disagree on the minimum macOS version. The README gives macOS Sequoia 15.1 as the minimum for the app, and Homebrew's record for the cask gives macOS 15, while the EXO Labs website says the app requires macOS Tahoe 26.2 or later.[1][40][41]

When Macs are cabled by Thunderbolt, the app offers to install a system service that creates a network location named exo, gives each Thunderbolt port its own network service and turns off Thunderbolt Bridge. This needs an administrator password, and the README documents how the app's uninstaller and a separate script reverse it.[1]

Published benchmarks

The figures below are as reported by the named source on the date given. They come from different software versions, models, quantization levels and prompts, and are not comparable with one another.

DateSourceHardwareModelReported result
2024-11-13VentureBeat, reporting EXO Labs [15]4 Mac Mini M4 and 1 MacBook Pro M4 MaxQwen2.5-Coder 32B18 tokens per second
2024-11-13VentureBeat, reporting EXO Labs [15]Same clusterNemotron 70B8 tokens per second
2024-12-25EXO Labs blog [36]1, 2 and 3 M4 Pro MacsLlama 3.2 3BOne request at a time: 49.3, 44.4 and 39.7 tokens per second. Several at once: 49.3, 95.7 and 108.8
2024-12-26EXO Labs blog [17]8 Mac Mini M4 Pro, 64 GB eachDeepSeek V3 671B, 4-bit5.37 tokens per second; 2.91 s to first token
2024-12-26EXO Labs blog [17]Same clusterLlama 3.1 405B, 4-bit0.88 tokens per second; 29.71 s to first token
2024-12-26EXO Labs blog [17]Same clusterLlama 3.3 70B, 4-bit3.89 tokens per second; 3.14 s to first token
2025-10-15EXO Labs blog [22]DGX Spark and M3 Ultra Mac StudioLlama 3.1 8B, FP16, 8,192-token prompt, 32 tokens generated2.32 s in total, against 6.42 s on the Mac Studio alone and 4.34 s on the DGX Spark alone
2025-12-18Jeff Geerling [39]4 Mac Studio M3 Ultra, RDMA over Thunderbolt 5Qwen3 235B32 tokens per second on the full cluster
2025-12-18Jeff Geerling [39]Same clusterKimi K2 Thinking"around 30 tokens per second"

Expanded article table

The Llama 3.2 result is the one EXO Labs used to explain the limits of pipeline sharding: a single request ran slower on three machines than on one, and only the combined throughput of several simultaneous requests rose with the number of devices.[36] The post on DeepSeek V3 attributed that model's higher speed than the much smaller Llama 3.3 70B to its mixture of experts design, in which only part of the parameters is used for each token.[17]

The October 2025 result came from running the prompt-processing (prefill) phase on the DGX Spark and the token-generation (decode) phase on the Mac Studio, while streaming the KV cache between them layer by layer.[22] The post's headline says "4x Faster LLM Inference"; its results report a 2.8 times overall speedup over the Mac Studio alone and describe the DGX Spark's prompt processing as 3.8 times faster than the Mac Studio's.[22] Tom's Hardware summarized it as a 2.8 times gain over the Mac Studio alone and noted that the build used was not yet public.[23]

The project's own sources disagree on whether a DGX Spark can be used with the public software. The company's home page names the DGX Spark beside the Mac Studio under the heading "Your devices acting as one.", while its answer on supported hardware names Apple silicon Macs and Linux installations from source.[41] The current README says "On Linux, exo currently runs on CPU" and that a Mac and a Linux machine do not join each other, although its Linux instructions mention the mlx-cuda13 and mlx-cuda12 extras for NVIDIA hardware.[1] The platform list, last changed in December 2025, places CUDA support for the DGX Spark under planned work.[8]

Geerling ran his tests on four Mac Studios lent by Apple: two with 512 GB of memory, which he priced at $11,699 each, and two with 256 GB at $8,099 each, a stack he described as costing "just shy of $40,000". He compared exo with llama.cpp's RPC mode on the same machines and reported that, with Qwen3 235B, exo sped up as nodes were added while llama.cpp slowed down.[39] The README reproduces three of his charts as its benchmarks section.[1]

Limitations and criticism

Single requests and networks

By EXO Labs' own 2024 account, a model that fits on one device runs a single request more slowly when spread over several by pipeline sharding, because each hop adds network delay.[36] The README ties its speed claims to tensor parallelism and to RDMA over Thunderbolt 5, and MLX's documentation calls the RDMA backend necessary for tensor parallelism.[1][37]

Cabling

The README's RDMA caveats are that every device must be cabled directly to every other device, that the cables must support Thunderbolt 5, that one Thunderbolt 5 port on the Mac Studio (the one beside the Ethernet port) cannot be used, and that all Macs must run exactly the same macOS version.[1] MLX's documentation likewise states that JACCL supports only fully connected topologies.[37] Geerling wrote that, as far as he could tell, Thunderbolt 5 switches did not exist, so every Mac had to be plugged into every other, and that he had seen RDMA clusters of only four Macs.[39]

Platform support

The first version advertised phones, Raspberry Pi boards and NVIDIA GPUs; the current documentation supports Apple silicon Macs and, from source, Linux on CPU.[3][1][8] All nodes must run the same version, and the README states that mixing versions in one cluster is not supported.[1]

Security of exposed nodes

The API reference describes no authentication mechanism, and the source code binds the API to all network interfaces (API_BIND_HOST = "0.0.0.0").[35][43]

On 2 March 2026 Kevin Breen, Senior Director of CTI at Immersive, published an account of a remote code execution flaw in exo. He wrote that a default installation had no authentication, listened on 0.0.0.0 and had a cross-origin policy open to any site, and that the code loaded a model's tokenizer with trust_remote_code hard-coded to true, so that adding a crafted Hugging Face model and placing an instance of it ran the attacker's code. According to the post, the developers released patches removing that default within a few days of his report, and the release did not change the cross-origin policy or the lack of authentication.[47] The current README says models requiring trust_remote_code "must be explicitly enabled (default is false)".[1]

On 2 April 2026 a user reported in issue #1833 that any host able to reach a node could list running model instances and delete them. The GitHub account Evanev7, which GitHub labels a collaborator on the repository, closed the issue the same day as not planned, with the reply: "this is intended behaviour, exo's api is unauthorized. run exo behind a reverse proxy and provide your own authentication if you wish to expose it downstream."[42] The report was later published as CVE-2026-50759 (2026-07-21). The National Vulnerability Database entry describes it as affecting exo 1.0.69 and lists a CVSS 3.1 base score of 7.5 (High).[48] A separate entry, CVE-2026-14738, published on 2026-07-05 from a VulDB submission, describes a weak hash in the function that builds cache keys for image features in versions up to 1.0.71, with a CVSS 4.0 score of 2.9 (Low), and says a pull request with a fix "awaits acceptance".[49]

The general risks of exposing a model server are covered in self-hosting LLMs.

Stability

Geerling, testing prerelease software in December 2025, wrote of RDMA over Thunderbolt: "When it works, it works great", and added that when it did not, he was glad to have automation set up to shut down and reboot the whole cluster quickly.[39] The notes for v1.0.68 in February 2026 said the release was meant to address "the stability issues users were running into on previous versions".[29]

Development in private

Geerling also wrote that he had "some underlying trust issues with Exo, since the developers went AWOL for a while", and that he wished the 1.0 version had not been developed in secrecy, while crediting the developers for releasing it under Apache 2.0.[39] A second user-opened issue, dated 3 August 2026, asked whether the repository was still being maintained; it was open on 19 November 2026.[44]

License

DateEvent
2024-07-16GNU GPL version 3 license file added to the repository [4]
2024-12-31EXO Labs blog announces a move to "a more open license" in early 2025 and asks people who have contributed code to join a Discord channel to coordinate the change [18]
2025-12-17Commit "prep repo for v1" removes the first version's files except the README, including the GPL license file; the repository exo-explore/ex-exo, now archived and carrying the GPL file, is created the same day [25][26]
2025-12-18Merge commit "exo v1" adds the rewritten code without a license file; an Apache License 2.0 file is added later that day, and its copyright line is then set to "Copyright 2025 Exo Technologies Ltd" [5][6][45]

Expanded article table

As of November 2026 the repository's license file is the Apache License 2.0, and GitHub's API reports the SPDX identifier Apache-2.0.[46][11] The GPL file accompanied the first version's code from 16 July 2024 until that code was removed from the main branch on 17 December 2025. It still accompanies that code in ex-exo and in the earlier commits of the main repository, where the license file at the last first-version commit is the GPL version 3.[4][25][26][51] The Apache file has accompanied the rewritten code since 18 December 2025.[6] The GPL is a copyleft license and Apache 2.0 a permissive one; the difference is explained in open-source license. The license covers exo's own code, not the models a cluster downloads.

EXO Labs

EXO Labs (written Exo Labs on its website and "exo labs" in the README) is the company that maintains exo.[41][1] VentureBeat described it in November 2024 as a startup founded in March 2024, named Alex Cheema and Mohamed "Mo" Baioumy as its co-founders, said Cheema had previously worked in distributed systems engineering, and reported that the company had raised an undisclosed amount from private investors.[15] heise online called it a small British company in November 2025.[2] The copyright notice in the repository's license names Exo Technologies Ltd.[45]

The company's website describes exo as "the open-source software that connects your devices into a local AI cluster" and says the company also owns and operates local.ai, a site about which models a given computer can run. It offers deployments to businesses through an inquiry form and lists Claude Code, OpenAI Codex, Hermes Agent and OpenCode as agent tools that can use a cluster.[41] Its blog has also covered training across consumer devices with DiLoCo, most recently in a post that the blog index lists with the date 15 September 2026 and describes as written in January 2026.[16]

Continue reading

More from See also

More to read

Topics:AI InferenceAI InfrastructureDeveloper ToolsOpen Source AI

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42exo labs. "exo: Run frontier AI locally" (README.md, main branch, last changed 2026-11-04). GitHub, accessed 2026-11-19. github.com/...README.md
  2. ^1 ^2 ^3Ben Schwan. "AI Cluster: Four Macs with 2 TB can run the giant model Kimi K2 Thinking". heise online, 2025-11-20. heise.de/...-giant-model-Kimi-K2-Thinking-11085716
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13exo labs. "exo: Run your own AI cluster at home with everyday devices" (README.md of the first version, at commit 5f18fae, 2025-10-30). GitHub. github.com/...README.md
  4. ^1 ^2 ^3 ^4 ^5exo-explore/exo. "add license" (commit bde1e53, adds the GNU General Public License version 3). GitHub, 2024-07-16. github.com/...bde1e53f5f
  5. ^1 ^2 ^3 ^4exo-explore/exo. "exo v1" (merge commit f64d17f). GitHub, 2025-12-18. github.com/...17fac01fba83e65fa258dd1bb9f9cce7ebee
  6. ^1 ^2 ^3 ^4 ^5exo-explore/exo. "re-add LICENSE" (commit 57d3813, adds the Apache License 2.0). GitHub, 2025-12-18. github.com/...57d3813692
  7. ^1 ^2 ^3 ^4 ^5exo-explore/exo. "AGENTS.md" (project overview and architecture). GitHub, accessed 2026-11-19. github.com/...AGENTS.md
  8. ^1 ^2 ^3 ^4 ^5 ^6exo-explore/exo. "EXO Platform support (partial roadmap)" (PLATFORMS.md, last changed 2025-12-18). GitHub, accessed 2026-11-19. github.com/...PLATFORMS.md
  9. ^Open Source Initiative. "Apache License, Version 2.0". opensource.org, accessed 2026-11-19. opensource.org/...apache-2-0
  10. ^Open Source Initiative. "GNU General Public License version 3". opensource.org, accessed 2026-11-19. opensource.org/...gpl-3-0
  11. ^1 ^2 ^3GitHub REST API. Repository record for exo-explore/exo (stars, forks, creation date, license), retrieved 2026-11-19. api.github.com/...exo
  12. ^1 ^2 ^3 ^4 ^5exo-explore/exo. "Releases". GitHub, accessed 2026-11-19. github.com/...releases
  13. ^1 ^2 ^3exo-explore/exo. "Serve requests without the master: each request streams over HTTP from the device that runs it (#2597)" (commit a60b956). GitHub, 2026-11-06. github.com/...a60b9569
  14. ^exo-explore/exo. "scaffolding for networking, inference and orchestration" (commit a21f59f, the repository's first commit). GitHub, 2024-06-23. github.com/...a21f59ff45
  15. ^1 ^2 ^3 ^4 ^5 ^6Carl Franzen. "You can now run the most powerful open source AI models locally on Mac M4 computers, thanks to Exo Labs". VentureBeat, 2024-11-13. venturebeat.com/...m4-computers-thanks-to-exo-labs
  16. ^1 ^2Exo Labs. "Blog" (index of posts with dates). blog.exolabs.net, accessed 2026-11-19. blog.exolabs.net
  17. ^1 ^2 ^3 ^4 ^5EXO Labs. "Running DeepSeek V3 671B on M4 Mac Mini Cluster - 12 days of EXO" (Day 2). blog.exolabs.net, 2024-12-26. blog.exolabs.net/day-2
  18. ^1 ^2EXO Labs. "2024 Wrapped - 12 days of EXO" (Day 7). blog.exolabs.net, 2024-12-31. blog.exolabs.net/day-7
  19. ^exo-explore/exo. "Merge pull request #737 from exo-explore/handlegzipdownload" (commit af734f1, tagged v0.0.15-alpha). GitHub, 2025-02-25. github.com/...af734f1bf6
  20. ^1 ^2exo-explore/exo. "Exo Project Status" (issue #819, opened 2025-04-14, closed 2025-12-18). GitHub. github.com/...819
  21. ^exo-explore/exo. "Update." (commit 56f783b, adds a notice to the README). GitHub, 2025-10-21. github.com/...83b38dc6b08ce606b07a5386dc40dae00330
  22. ^1 ^2 ^3 ^4EXO Labs. "NVIDIA DGX Spark™ + Apple Mac Studio = 4x Faster LLM Inference with EXO 1.0". blog.exolabs.net, 2025-10-15. blog.exolabs.net/nvidia-dgx-spark
  23. ^1 ^2Zak Killian. "Two Nvidia DGX Spark systems fused with M3 Ultra Mac Studio to deliver 2.8x gain in AI benchmarks: EXO Labs demonstrates disaggregated AI inference serving". Tom's Hardware, 2025-10-16. tomshardware.com/...achieves-a-2-8-benchmark-boost
  24. ^Devindra Hardawar. "You can turn a cluster of Macs into an AI supercomputer in macOS Tahoe 26.2". Engadget, 2025-11-18. engadget.com/...puter-in-macos-tahoe-262-191500778
  25. ^1 ^2 ^3exo-explore/exo. "prep repo for v1" (commit 0fcee70). GitHub, 2025-12-17. github.com/...e70833a9d058881711543d9c04a1a60c7d02
  26. ^1 ^2 ^3exo-explore. "ex-exo" (archived repository of the first version, GPL-3.0, created 2025-12-17). GitHub, accessed 2026-11-19. github.com/...ex-exo
  27. ^exo-explore/exo. "v1.0.61" (release notes). GitHub, 2026-01-08. github.com/...v1.0.61
  28. ^1 ^2exo-explore/exo. "EXO v1.0.63 Release Notes". GitHub, 2026-01-16. github.com/...v1.0.63
  29. ^1 ^2 ^3 ^4 ^5exo-explore/exo. "EXO v1.0.68 Release Notes". GitHub, 2026-02-25. github.com/...v1.0.68
  30. ^exo-explore/exo. "EXO v1.0.69 Release Notes". GitHub, 2026-03-27. github.com/...v1.0.69
  31. ^exo-explore/exo. "EXO v1.0.70 Release Notes". GitHub, 2026-04-17. github.com/...v1.0.70
  32. ^exo-explore/exo. "EXO v1.0.71 Release Notes". GitHub, 2026-04-23. github.com/...v1.0.71
  33. ^1 ^2 ^3exo-explore/exo. "libp2p -> zenoh (#2132)" (commit 09f9ea3). GitHub, 2026-06-03. github.com/...09f9ea313f
  34. ^1 ^2exo-explore/exo. "pyproject.toml" (package version, Python requirement and MLX extras). GitHub, accessed 2026-11-19. github.com/...pyproject.toml
  35. ^1 ^2 ^3 ^4 ^5exo-explore/exo. "EXO API - Technical Reference" (docs/api.md, last changed 2026-11-03). GitHub, accessed 2026-11-19. github.com/...api.md
  36. ^1 ^2 ^3 ^4EXO Labs. "Transparent Benchmarks - 12 days of EXO" (Day 1). blog.exolabs.net, 2024-12-25. blog.exolabs.net/day-1
  37. ^1 ^2 ^3 ^4 ^5MLX contributors. "Distributed Communication". MLX documentation, accessed 2026-11-19. ml-explore.github.io/...distributed
  38. ^Apple. "macOS Tahoe 26.2 Release Notes". Apple Developer Documentation, accessed 2026-11-19. developer.apple.com/...macos-26_2-release-notes
  39. ^1 ^2 ^3 ^4 ^5 ^6 ^7Jeff Geerling. "1.5 TB of VRAM on Mac Studio - RDMA over Thunderbolt 5". jeffgeerling.com, 2025-12-18. jeffgeerling.com/...studio-rdma-over-thunderbolt-5
  40. ^1 ^2Homebrew. Cask record for "exo" (version 1.0.71, requires macOS 15 or later on Apple silicon). Homebrew Formulae API, retrieved 2026-11-19. formulae.brew.sh/...exo.json
  41. ^1 ^2 ^3 ^4 ^5Exo Labs. "Exo Labs: Frontier models. On hardware you own." (company home page). exolabs.net, accessed 2026-11-19. exolabs.net
  42. ^exo-explore/exo. "Unauthenticated Remote Instance Termination" (issue #1833, opened and closed 2026-04-02). GitHub. github.com/...1833
  43. ^exo-explore/exo. "src/exo/api/listening_socket.py" (defines API_BIND_HOST). GitHub, accessed 2026-11-19. github.com/...listening_socket.py
  44. ^exo-explore/exo. "Is this repository even actively being watched and maintained in the last 4 months?" (issue #2258, opened 2026-08-03). GitHub, accessed 2026-11-19. github.com/...2258
  45. ^1 ^2 ^3exo-explore/exo. "LICENSE: Fix company name/year" (commit 9e2bdee). GitHub, 2025-12-18. github.com/...9e2bdeef92
  46. ^exo-explore/exo. "LICENSE" (Apache License, Version 2.0). GitHub, accessed 2026-11-19. github.com/...LICENSE
  47. ^Kevin Breen. "Unauthenticated RCE in EXO: Why the Security Architecture of Open-Source AI Platforms Needs Immediate Attention". Immersive, 2026-03-02. immersivelabs.com/...rms-needs-immediate-attention
  48. ^National Institute of Standards and Technology. "CVE-2026-50759". National Vulnerability Database, published 2026-07-21. nvd.nist.gov/...CVE-2026-50759
  49. ^National Institute of Standards and Technology. "CVE-2026-14738". National Vulnerability Database, published 2026-07-05, last modified 2026-10-25. nvd.nist.gov/...CVE-2026-14738
  50. ^exo-explore/exo. Commit history of the main branch, 4 June to 31 October 2026. GitHub, accessed 2026-11-19. github.com/...main
  51. ^exo-explore/exo. "LICENSE" (GNU General Public License version 3, at commit 5f18fae, 2025-10-30). GitHub. github.com/...LICENSE

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

1 revision · v1 · 5,200 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Independent fact-check before publication on November 19, 2026: 168 claims checked against 58 cited and independent sources; 2 material and 10 minor corrections applied before publication.

Cite this page: AI Wiki. "exo (software)." aiwiki.ai, updated 19 Nov 2026, fact-checked 19 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/exo

Edit article

What links here