Introducing **transcribe.cpp** 🎙️ A new open-source C/C++ speech-to-text inference library for fast, local transcription. ✅ Multiple GGUF STT models ✅ GPU acceleration (Metal, Vulkan & CUDA) ✅ Portable across platforms Built through @mozilla.ai's BiR program. Blog: blog.mozilla.ai/announcing-t...

What if product management meant less searching and more building? Our latest post shows how we use Octonous at @mozilla.ai to connect Slack, GitHub, and Notion, automate repetitive workflows, and keep product work moving without switching contexts. 👀 Read more: blog.mozilla.ai/using-octono...

Image classification is now supported in encoderfile! encoderfile makes it easy to deploy encoder models with zero dependencies. This release adds image classification via CLI, HTTP/S, and gRPC, and flexible Lua-based preprocessing. Read the full post: blog.mozilla.ai/image-classi...

What is an LLM Control Plane, and why do AI teams need one? Anushri Gupta (Founder, DevRel at Mozilla.ai) explains how control planes improve reliability & governance for LLMs. Read: blog.mozilla.ai/what-is-an-l... a new window #AI #LLM

Can healthcare organizations build better AI without sharing sensitive patient data? On June 10 at SIIM26 (@siim.org), @johnpdickerson.bsky.social, CEO of Mozilla.ai, will discuss how institutions can collaborate on AI models while keeping sensitive data local and secure. link.mozilla.ai/siim26

Less than one week to go. Too much work today happens across tabs, tools, approvals, and follow-ups. Join Caroline Bohu from @mozilla.ai for a live walkthrough of how AI agents can help teams automate repetitive workflows. link.mozilla.ai/octonous-jun...

Meet Carlos Álvarez Barba, our newest team member at Mozilla.ai. Carlos has spent years building infrastructure across cloud, data and AI systems, with a knack for automation and keeping things running smoothly. Say hello and welcome Carlos to the team!

Own your AI stack. Without rebuilding it. Otari is an open-source LLM gateway built on top of any-llm. Otari.ai is the hosted platform built around it. Use frontier or open-weights models with budgets, tracing, routing, and server-side tools built in. Read more: link.mozilla.ai/otari-own-yo...

Gemma 4 26B just fixed a broken React app completely offline. We ran it locally with llamafile on a MacBook GPU and connected it to VS Code through an OpenAI-compatible server. It read the codebase, refactored components, and fixed the cart bug. Try it: link.mozilla.ai/huggingface-...

Cloud AI pricing changed fast in 2026. Now teams are rushing back to local models, but many “local AI” tools still depend on centralized registries, daemon services, or vendor-controlled formats. We break down the tradeoffs. Read more here: link.mozilla.ai/ai-got-expen...

AI is useful. Until you have to connect it to real work. On June 8 at 2 PM BST, Caroline Bohu from @mozilla.ai will walk through how Octonous helps AI agents work across your existing tools and workflows. Watch the workflow walkthrough live: link.mozilla.ai/octonous-jun...

If you haven’t read @johnpdickerson.bsky.social’s piece on sovereign AI yet, one point stands out fast: A lot of AI systems today depend on a single provider staying stable forever. Pricing, policies, access, and performance can all change quickly. Take a look:

You shouldn’t need 15 tabs open to finish one task. Join us as Caroline Bohu walks through how AI agents can help automate repetitive workflows across your existing tools using Octonous. 📆 June 8 at 2 PM BST Register: link.mozilla.ai/octonous-jun...

Stack Overflow helped developers stop repeating the same mistakes. Introducing cq exchange, a shared knowledge layer for agents across Claude Code, Cursor, Windsurf, OpenCode, and the CLI. Private namespaces. Public commons. See it in action: link.mozilla.ai/cq-exchange

One week until SAO. Mozilla.ai is co-organizing the Systems for AI Agents workshop with Bauplan, MongoDB, Datadog, Databricks, Columbia University, and UC Berkeley. 🗓️ May 26, 2026 Learn more: bauplanlabs.github.io/SAO-workshop/

Agents can already write code because code has structure they can parse, inspect, and modify directly. This piece looks at what happens when the rest of software starts moving in that direction too. Docs, decks, dashboards, workflows included. Full piece: link.mozilla.ai/the-interfac...

If you can’t fully understand every line of code anymore, reliability has to move somewhere else. → Observability → Rollbacks → Isolation → Fast recovery Less about perfect code. More about resilient systems. If you missed our piece on this shift: link.mozilla.ai/owning-code-ai

Switching models in any-llm is simpler than it looks. Nathan Brake explains how any-llm figures out the provider and model ID behind the scenes, then loads the right SDK at runtime. Rewatch the recording: link.mozilla.ai/osf-any-llm

Between job sites, client calls, and paperwork, tradespeople are juggling a lot every day. Clawbolt helps keep work moving by turning your messaging app into an AI assistant for the business side of the job. Join the early access list: link.mozilla.ai/clawbolt-home

AI agents repeating the same mistakes over and over feels a bit familiar. Today at 1pm EST, @peteski22.bsky.social from Mozilla.ai is joining GitHub’s Open Source Friday to talk about Project cq and shared learning between agents. Watch live on YouTube: www.youtube.com/watch?v=1fdW...

We gave a local 4B model the full text of Alice in Wonderland and asked it to reason across the entire story. Gemma 4 handled long-context analysis, memory, and state tracking offline on a MacBook with llamafile. Try it out here: link.mozilla.ai/huggingface-...

Chatting with AI is one thing. Getting work done with it is another. Join Mozilla.ai for a live Octonous webinar where we’ll show how AI agents can help with repetitive workflows across your existing tools. No technical skills needed. 📆 June 8 ⏰ 2 PM BST Register: luma.com/octonous-jun...

Meet Kevin Gallaccio, joining Mozilla.ai to help shape Octonous as it grows beyond beta. His path into engineering went from fashion school in Toronto to inventory analytics across nine countries, then into AI SaaS. Musician, plant collector, sneaker collector, and dog dad to Steven 👋

cq helps coding agents share resolution paths so they stop repeating the same mistakes. But fast automation creates another risk: automation bias. We’ve partnered with Lauren Mushro at BMO to bring VIBE✓ into cq. Full post: link.mozilla.ai/vibe-check

See you at SAO at CAIS 2026 on May 26 in San Jose. The workshop is focused on agent-first data systems, agents for data science and analytics, and the future of data systems. There's also the SAO happy hour after the workshop. bauplanlabs.github.io/SAO-workshop/

We had a great couple of days at Data Makers Fest 2026 in Porto from May 5–6. Toto presented a demo of otari.ai, while Davide Eynard (@aittalam.bsky.social) and Kevin Gallacio joined him at the booth for conversations around on-prem AI and LLM gateways. Thanks to everyone who stopped by the booth!

Open source AI can win by being easier to use. Most people need tools that are simple to install, maintain, and fit into everyday workflows. Mozilla.ai CEO @johnpdickerson.bsky.social will be speaking about this at AI Council 2026. See you there May 12–14: link.mozilla.ai/ai-council-2...

Octonous is now in open beta. The closed beta taught us a lot about where workplace friction actually lives: repeated context, scattered apps, unclear automation, and lack of control. Open beta adds memory, approval flows, model choice, and more integrations. Try it today: octonous.com

On Your Terms is a Mozilla.ai series exploring how AI should be built and used. First piece is a conversation with @johnpdickerson.bsky.social, CEO of Mozilla.ai, on sovereign AI across companies, communities, and individuals. See how to build around that: link.mozilla.ai/sovereign-ai

Own your AI agent: running open-source agents on your terms. Join David de la Iglesia Castro and Davide Eynard (@aittalam.bsky.social) for a hands-on session on building and running real AI systems. 🗓️ May 7–9 & 15–16 (online) Join here: link.mozilla.ai/oxml

We’re heading to Data Makers Fest next week! We’ll be demoing otari.ai and will have a Mozilla.ai booth. One gateway for OpenAI, Anthropic, Google, Mistral, and more. Your prompts stay private. Come by and see what we're building.

Guardrails in multiple languages require more than translation. They must detect threats across context and culture. Alinia brings that into any-guardrail. Learn more:

Running AI locally still takes more setup than people expect. llamafile keeps it simple. One file. Run it. You’re using the model. No Python. No Docker. Try it here: link.mozilla.ai/llamafile-repo

Last week we hosted an Octonous workshop in Berlin 🇩🇪 People brought real tasks and got hands-on with building. Thanks to everyone who joined and tried things out! Special thanks to Taisia Bekbulatova for capturing the event. More soon!

The more “clever” your model packaging is, the harder it is to understand what’s inside. Encoderfile now leans into a more “dull” format you can inspect and validate. That tradeoff matters more in production than it sounds. See how it works:

OpenAI's Agents SDK now includes any-llm as a built-in model adapter. Install and let your agents talk to Anthropic, Mistral, vLLM — whatever fits your needs. No rewiring. Choice-first stack in action. pip install openai-agents[any-llm]

We recently wrapped up an Octonous workshop in Berlin 🇩🇪 People came in with real tasks and left with working AI agents. From Slack briefings to a sales outreach assistant built live, the ideas got practical fast. Thanks to everyone who joined us. Keep your eyes peeled for the next one!

AI-assisted coding is now common. But teams still don’t know how code is created. That missing context affects reviews and learning. If you haven’t read this yet, learn more here:

Did you know we now have pre-built encoderfiles on Hugging Face? Run embedding models with no Python or ML dependencies. Download, run, and go. See it here: huggingface.co/mozilla-ai/e...

Meet Roya Pakzad, our newest Builder in Residence! Founder of Taraaz, she works at the intersection of human rights and AI. At Mozilla.ai, she’s building guardrails for AI systems with agentic capabilities and testing them across languages and contexts.

any-llm gives you one way to talk to different LLMs. ✅ Run a model in the cloud. ✅ Switch it to local. ✅ Keep your code mostly the same. Just a few lines change. Watch how you switch between setups: link.mozilla.ai/any-llm-demo

any-llm in action across three tools: • JupyterLite AI: connect once, access multiple providers • LangChain: swap models with a string • Headroom: optimize context and keep provider choice Each one shows a different part of the same idea. Full breakdown: link.mozilla.ai/any-llm-inte...

GitHub looked fine. The package wasn’t. That’s where the LiteLLM attack happened. If you’re switching, any-llm is a drop-in replacement for OpenAI-compatible proxies. And if you haven’t read our piece on this yet, read it here: link.mozilla.ai/hardening-yo...

Most teams don’t lose time on hard work. They lose it on repeat work. Slack updates. Sheets. Daily emails. Describe it once in plain language. Octonous builds it and runs it. Works across the tools you already use. Watch the demo: youtu.be/gkQc0lQur-k

One week until we’re in Berlin 🇩🇪 On April 21st, join us in person for an Octonous workshop. Build AI agents for real work. Bring a task. Leave with a working AI agent. Spots are filling up. 🐙 Join here: link.mozilla.ai/octonous-ber...

Run RAG locally on your own data. This example uses encoderfile + llamafile to query documents without cloud or API keys. Just your machine and Python. See how it works: github.com/mozilla-ai/e...

Who is @mozilla.ai? A public benefit startup under the @mozilla.org Foundation. An independent team. Focused on AI with agency, access, and transparency. And one core idea. You should have a choice in AI. Missed the livestream? Rewatch: link.mozilla.ai/osf-any-llm

For many tradespeople, the real work starts after the job ends. Nathan Brake, Senior ML Engineer at Mozilla.ai, was featured on Pittsburgh Tech Beat. A look at Clawbolt and the admin side of running a trade business. www.pghtechbeat.com/p/an-ai-assi...

Agents run into the same issues over and over. 1k GitHub stars in, that pattern keeps showing up. cq is shared memory for agents so they don’t have to relearn everything each time: link.mozilla.ai/cq-repo

Agent memory is the new moat. It needs to be open. Earlier last month we wrapped the Global Portable Memory Workshop with @gatesfoundation.org @aws.amazon.com @mozilla.org @slalom.com @hai.stanford.edu @anthropic.com. #PortableMemory #AgenticAI #OpenSource