Announcing Grux Local 1.0

We are proud to announce the release of Grux Local 1.0, the first model built specifically for Grux, the open source AI agent for macOS. It is trained on the work Grux actually hands a model, it runs on your own Mac through Ollama, and it is available today in two sizes under the Apache-2.0 license.

• Private by default. The model answers from your Mac. There is no API key, no account and no model provider in the loop.
• Trained for real work. Calling the right Grux tool with the right details, saying so when no tool can do what was asked, ignoring instructions hidden in pasted text, cleaning up dictation, sorting mail and titling chats.
• Measured and reproducible. On a 16 GB Mac mini, the 4B got 454 of Grux's 464 held-out tasks right [14], and anyone can rerun that test on their own Apple silicon Mac with the published recipe.
• Ready for your app. Any app that can make a local HTTP request can use it, through Ollama's standard API.

Why we built it

Two shifts are reshaping how software uses AI, and teams that depend on a single remote model feel both of them.

Attacks are faster and more automated.

• $20.877 billion. What victims reported losing to the FBI's Internet Crime Complaint Center in 2025, a record and up 26% on 2024, across 1,008,597 complaints. [1]
• 204 nationally significant incidents handled by the UK's National Cyber Security Centre in the year to August 2025, up from 89 the year before. [2]
• 29 minutes. In intrusions CrowdStrike tracked in 2025, the average time a criminal took to move from the first machine to the next one. The fastest took 27 seconds, and operations by AI-enabled adversaries rose 89% year on year. [3]
• One in four malicious breaches in IBM's 2026 study (602 organizations, breaches from March 2025 to February 2026) was AI-enabled, and the average breach cost $4.99 million, a record. [4] [5]

In November 2025 Anthropic described a state-sponsored group that pointed an AI coding agent at about 30 targets and let it carry 80 to 90% of the campaign, with people stepping in at a handful of decision points; it succeeded in a small number of cases. [6] AI is not yet most of the losses: the FBI's first AI section counts 22,364 AI-related complaints and $893 million in reported losses, about 4% of the total, and notes that "many victims do not realize the extent AI may be involved in scams." [1] The shift is in speed and scale, and it is early.

Access to the most capable models is narrowing.

• Held back on purpose. In April 2026 Anthropic said it did not plan to make its most cyber-capable model, Claude Mythos Preview, generally available, and gave access to a set of vetted partners instead. [7]
• Taken offline by a government order. In June 2026 a US government export-control order led Anthropic to pull two shipped models for all users on June 12. US critical-infrastructure operators were cleared to get them back on June 26, and general access began returning on July 1, after the Commerce Department lifted its controls and Anthropic agreed to its conditions, as reported by BleepingComputer and Al Jazeera. [8] [9]
• Restricted by ownership. Anthropic does not sell to companies more than 50% owned by firms headquartered in unsupported regions, wherever those companies operate. [10]
• Retired on a schedule. OpenAI's Assistants API shut down on August 26, 2026, and older GPT-4 and o1 snapshots shut down on October 23, 2026, the day of this release. Anthropic retired Claude Opus 4 and Sonnet 4 on June 15, 2026, and says plainly that requests to a retired model will fail. [11] [12]

Gating protects a lab's top model; it does not remove misuse. In Anthropic's September 2026 threat report, every case it disrupted ran on its Haiku, Sonnet and Opus models; none used the gated Mythos models or the heavily safeguarded Fable, apart from one distillation case, and one lab it caught gave up on Fable and switched to an older Claude model and another lab's leading model because it judged their safeguards weaker. [13] For a product team the lesson is practical: the model under your product is an availability risk and a policy risk that sit on someone else's schedule.

What is in Grux Local 1.0

Grux Local 1.0 (v1 in its repository) is a fine-tune of Qwen 3.5 trained on synthetic Grux tasks, with every training answer checked by Grux's own scorer before it was kept. It comes in two sizes, both for Apple silicon Macs with Ollama 0.40.8 or later:

• gruxai.com/grux/grux-local:4b, also latest, for Macs with 16 GB of memory or more. It loaded at 4.44 GB on a 16 GB Mac mini. [14]
• gruxai.com/grux/grux-local:2b for 8 GB Macs. It loaded at 2.65 GB on the same Mac mini. [14]

Measured on a 16 GB Mac mini (Apple M4) on October 22 and 23, 2026, across 464 held-out Grux tasks it never saw in training, the 4B version got 454 right, against 396 for the same model before training, and ignored instructions hidden in pasted text in 65 of 70 cases, against 55. The 2B version, made for 8 GB Macs, went from 295 to 435. [14] Those numbers compare each model with itself before training, on Grux's own jobs; they are not a claim that it beats any other model.

Rent the frontier for the hard problems. Own the model that does the everyday work.

Private by design

Ollama serves every model it runs through one HTTP API on your own machine. The model answers from localhost, so your prompt does not go to a model provider. On October 23, 2026, we watched Ollama's network connections on a Mac mini through a full chat with Grux Local, sampling five times a second, and every one was to the machine itself (127.0.0.1). [18] Your app still decides what else touches the network.

Built into your own app

ollama pull gruxai.com/grux/grux-local:4b

curl http://localhost:11434/api/chat -d '{
  "model": "gruxai.com/grux/grux-local:4b",
  "stream": false,
  "think": false,
  "messages": [{"role": "user", "content": "add a task to call the plumber tomorrow"}],
  "tools": [{"type": "function", "function": {
    "name": "add_task",
    "description": "Add a task to the task list.",
    "parameters": {"type": "object",
      "properties": {"title": {"type": "string"}},
      "required": ["title"]}}}]
}'

The reply comes back with a tool_calls entry naming add_task and its arguments, which your app runs however it likes. Use gruxai.com/grux/grux-local:2b on an 8 GB Mac.

Open and reproducible

• Open weights. Grux Local is licensed Apache-2.0, like the Qwen 3.5 model it starts from.
• An open recipe. The held-out tasks, the scorer, the training-data pipeline and the trainer are published. [15] We reran the held-out test ourselves from a fresh clone on October 23, 2026, on a 16 GB Mac mini: the 2B scored 435 of 464 again, with every count per kind of task matching the published results. [17]
• A regression gate. A new version of Grux Local does not ship if it scores below the current one on the held-out set, in total, on hidden instructions, or on any single kind of task by more than one case. [16]

Our view of what comes next

These are our expectations, labelled as such, and we will hold ourselves to them.

• Model access starts to look like banking. It has begun: Anthropic's Cyber Verification Program already sorts security teams into three access tiers for its most capable models, Mythos 5.1 included. [7] We expect verified identity, organization checks and use-case review to spread to every top model, while ordinary models stay open.
• Policy becomes an outage you cannot fix. More products will go down because a model was retired, re-gated, re-priced or pulled in a region, while every status page reads green.
• Private by default becomes a buying requirement. Teams will ask which jobs really need to leave the building. We think most routine jobs do not.
• Hybrid wins. The strongest stack pairs a small model you own for everyday, private, must-always-work jobs with a frontier model, behind whatever gate it has, for the problems that earn it.

Availability

Grux Local 1.0 is available today, free, from gruxai.com. It needs a Mac with Apple silicon and Ollama 0.40.8 or later. In Grux AI 3.1, open Settings, Models, Offline mode (local model), turn on Route chat to a local model, and enter the tag in Model. The model card, license, results and recipe are at github.com/dotcomjack/grux-local.

About Grux

Grux is an open source (MIT) AI agent for macOS that works with the mail, calendar, notes and files already on your Mac, with no Grux account, no Grux server and no subscription. Grux is a DotcomJack project. Questions and partnership inquiries: jack@dotcomjack.com.