mastodon 4.8.0-nightly.2026-10-06Open, transparent AI for real world impact. Built for developers, creators, and teams shaping what’s next.
Agents shouldn’t have to figure everything out from scratch.
Right now, they do.
cq is a Stack Overflow for agents, where knowledge is shared and improved over time.
Less repetition. More reliable outcomes.
See how it works: https://link.mozilla.ai/cq-stack-overflow-for-agents
Local, private, opensource AI should be easy to use.
We just released llamafile 0.10.0, which brings llama.cpp’s latest models and features - such as multimodal and tool calling support - into a single, portable executable that runs on your hardware.
Explore the update: https://link.mozilla.ai/llamafile-v010
Running a transformer model usually starts with:
👎Large Docker images
👎Python dependencies
👎Slow setup
This demo shows a simpler way to run the same model.
👍 Single executable.
👍 Smaller footprint.
We also compare: 1.7GB vs 295MB.
Watch here: http://link.mozilla.ai/encoderfile-demo
Meet Ram M. Kripa, one of our newest Builders in Residence.
He builds AI agents that interact with the web like real users.
At Mozilla.ai, he’s exploring where browser-based agents can help when APIs fall short, and how to use them safely in real systems.
Agent memory is the new moat and it needs to be open.
Earlier last month we wrapped two days at the Global Portable Memory Workshop in Seattle. The room was full of people who get it: if portable, user-controlled memory consolidates inside closed platforms, users lose. We need open infrastructure for the AI era, the same way Linux and open protocols built the web.
Grateful to Gates Foundation, @mozilla@mastodon.social, Anthropic, and the full crew.
Meet CJ, one of our newest Builders in Residence.
He’s contributed to llamafile, created whisperfile, and built Handy.
At Mozilla.ai, he’s focused on exploring new ways to distribute and run local AI.
llamafile v0.10.0 includes:
• Multimodal support
• Tool calling + Anthropic Messages API
• CLI, HTTP server, terminal chat
• CPU + GPU optimizations
• Runs across different operating systems
All bundled into one executable.
If you missed the blog post, read it here: https://link.mozilla.ai/llamafile-v010
@fosdem@fosstodon.org 2026 kicks off this weekend (Jan 31–Feb 1) in Brussels.
If you’re attending, come visit the Mozilla booth and say hello.
@totostache@mastodon.social, Nathan Brake, and @mala@fosstodon.org will be there throughout the event.
See you at FOSDEM!
What is any-llm?
One interface to talk to any model.
Across local and self-hosted setups.
Missed the livestream? Watch the recording: https://link.mozilla.ai/osf-any-llm
What’s one task you repeat every week?
On April 21 in Berlin 🇩🇪, we’re hosting an in-person Octonous session.
Describe your task in plain language. Leave with a working AI agent.
Register here: https://link.mozilla.ai/octonous-berlin
Running a trade business includes a constant flow of admin work.
Quotes, scheduling, invoices, payments, reviews.
This work often continues after hours.
We shared how this gap shows up in the trades and what we’re building with Clawbolt to better support it.
Take a look: https://link.mozilla.ai/clawbolt
Julie, VP of Product at Mozilla.ai, was in Paris last week attending a Current AI event focused on AI in the public interest.
Led by Ayah Bdeir and her team, who we’ve worked with before.
Strong mix of conversations, people, and creative experimentation.
Good to see more efforts like this coming together.
KubeCon 2026 is underway in Amsterdam.
https://Mozilla.ai engineers are there this week, including Mario David Cariñana Abasolo, Peter Wilson, Dimitris Poulopoulos, and @totostache@mastodon.social.
If you are attending, come say hello.
Spent time at SCALE 23x connecting with people across the open source AI space.
Lots of booth traffic, strong interest in local models, and good conversations.
Thanks to everyone who stopped by!
We spent last week in Lisbon 🇵🇹 building AI agents with Octonous.
We brought together a small group to try it on real work.
People were focused, testing ideas, and exploring their own workflows.
One participant spent the session organizing their GitHub with a smile.
Another left with literal shiny eyes.
One is now continuing it as an ongoing experiment.
Keep your eyes peeled. We have more to come!
We’re happy to welcome Caroline Bohu to Mozilla.ai as our new Solutions Engineer.
Caroline focuses on making technology easier for people to understand and use, working closely with partners and teams adopting AI tools.
Outside of work she’s a proud mum who enjoys good food, great adventures, and a good book.
https://Mozilla.ai is now a launch partner of Flower Hub.
We published a new project called **fed-phish-guard**, which trains a phishing URL classifier using federated learning.
Each client trains locally and only sends model updates back to the server, keeping browsing data private.
Read the blog post here: https://link.mozilla.ai/flower-hub
Check the project page here: https://flower.ai/apps/mozilla-ai/fed-phish-guard/
Running a trade business means more than just doing the job.
Nathan Brake, Senior ML Engineer at https://Mozilla.ai, was featured on Pittsburgh Tech Beat.
The article highlights the admin work that fills nights and weekends, and how Clawbolt is being built to help manage it.
https://www.pghtechbeat.com/p/an-ai-assistant-for-the-trades
Agents run into the same issues over and over.
1k GitHub stars in, that’s still the pattern we see.
cq is shared memory so they don’t have to figure things out from scratch each time.
Open source, still early: https://link.mozilla.ai/cq-repo
Encoder models don’t chat, but they power most production NLP.
Deploying one still means dragging along a Python runtime, a full dependency tree, and serving infrastructure designed for something much larger.
Encoderfile takes a different approach.
A pre-built executable with an appended payload containing model weights, tokenizer data, and a manifest. At runtime, it reads itself and loads everything accordingly.
Code is becoming abundant.
Understanding and reliability are now the constraints.
That shift changes what engineering looks like.
If you missed our post on this shift: https://link.mozilla.ai/owning-code-ai
Since joining Flower Hub as a launch partner, we shared `fed-phish-guard`, a federated phishing URL classifier where training happens locally and only model updates are shared.
If you missed the blog post, you can find it here: https://link.mozilla.ai/flower-hub
There’s growing momentum around federated learning in real-world use cases.
We’ll also be at the Flower AI Summit. See you there: https://luma.com/fyjtra6v
We’re partnering with Data Makers Fest, a 2-day event for AI, ML, and data (May 5–6, 2026).
A chance to hear how people are actually building and working with data today.
Use code MOZILLA-10 for 10% off: https://www.datamakersfest.com/
Daniel Nissani (AI Safety Engineer) and Roya Pakzad (Senior Mozilla Fellow) evaluated three guardrails on multilingual humanitarian prompts:
• FlowJudge
• Glider
• AnyLLM (GPT-5-nano)
Each reviewed the same asylum-related scenarios in English and Farsi.
The LiteLLM incident wasn’t just about a compromised package.
It showed how easy it is to trust that source code and distributed packages match.
In this case, they didn’t.
If your LLM gateway manages your credentials, it’s worth reviewing how you handle dependencies.
What to check: https://link.mozilla.ai/hardening-your-llm-dependency-supply-chain
When building software becomes easy, the challenge shifts.
Velocity can quietly become the metric, even if nobody says it out loud.
Teams start shipping what’s easy to release instead of what actually matters.
When shipping becomes frictionless, product judgment becomes the real constraint.
See the full post: https://link.mozilla.ai/shipping-becomes-easy
We’re hosting the Octonous workshop in Lisbon 🇵🇹 today.
It’s a hands-on session where you’ll build an AI agent for a real task.
Bring a task and leave with a working AI agent.
No technical skills needed.
A few seats are still available.
Join here: https://link.mozilla.ai/octonous-lisbon