Welcome, humans.

What if you could run frontier-level AI models on your own hardware, without sending a single prompt to OpenAI's servers?

Sound impossible to you? Well, we're doing it live. Like, right now.

🔴 Join us RIGHT NOW at this week's Neuron Live as we run multiple open-weight AI models in real-time, walk through exactly how to get started, and explore what “open” actually means in AI (spoiler: there's a big difference between open-source and open-weight that matters for your privacy and control).

Plus, we’ll have a guest at 11am PST: Ece Kummar, CVP and Lab Director of AI Frontiers.

Here's what you'll see:

• Live demos running models with LM Studio and Hugging Face Spaces—no theory, just hands-on walkthroughs you can replicate.

• Microsoft Research's Fara-7B model—a special guest appearance from Ece Kamar (CVP & Managing Director of AI Frontiers Lab) to discuss Microsoft's approach to open-weight AI (she joins @ 11:00am PST, 1pm CT, 2pm ET).

• Real talk on costs, hardware requirements, and tradeoffs—what it actually takes to run these models outside closed platforms.

• The models worth trying today—we'll highlight the most interesting options available to anyone right now.

Why join live? Because you can ask questions in real-time as we troubleshoot, deploy, and test these models. Plus, if you've ever been curious about running AI without platform lock-in, this is your crash course.

Watch Live Now on YouTube

P.S. P.S. If you've been following the open Chinese AI wars, the new battle is between Kimi K2.5, and GLM 4.7.

Both Kimi K2.5 and GLM 4.7 are now achieving near Claude Sonnet 4.5-level coding performance on the benchmarks that actually matter.

On SWE-bench Verified, which tests whether AI can fix real bugs in real GitHub repositories, Kimi K2.5 scored 73.8%, meaning it successfully debugged 3 out of 4 real-world issues.

On LiveCodeBench—which tests coding ability on fresh programming problems the models couldn't have memorized—Kimi hit 85% while GLM 4.7 scored 84.9%, ranking #1 among open-source models.

Now, Microsoft's Fara-7B is playing a different game entirely.

It's a tiny 7B parameter model designed for computer use: meaning it can actually control a browser like a human, clicking buttons, filling forms, and navigating websites autonomously. On WebVoyager (which tests multi-step web navigation tasks), Fara-7B outperformed GPT-4o and can run entirely on-device with no cloud connection.

Meanwhile, Arcee's Trinity models are one of the first fully US-trained open-weight alternatives built from scratch. This live stream will show you how to run these types of models safely on your own hardware (or over the cloud for something like Kimi 2.5, which is massive, if we have time!)

P.P.S: We’ll be hanging out for the next two hours, so come through and ask your questions!

Keep scrolling for four of our most recent episodes covering coding agents, building agentic workflows, AI safety research, and how to think about AI as normal technology.

IN CASE YOU MISSED IT…

Recent podcast episodes worth your time:

This AI Agent Builds Better Code Than Most Developers (w/ Factory AI's Eno Reyes)

Watch on YouTube | Listen on Spotify | Apple Podcasts

Factory AI's Droids can handle 80 million token coding sessions (yes, really). We break down how they outperformed OpenAI and Anthropic in context compression, why Stanford found codebase quality is the ONLY predictor of AI success, and (45:17) the framework for making anything "agent-ready."

Why We MUST Monitor AI's Thinking (w/ OpenAI's Bowen Baker)

OpenAI researcher Bowen Baker discusses how reasoning models like o1 can plan and deliberate before answering, which introduces new safety risks (and new opportunities to monitor them). The episode explores “reward hacking” (when AI models cheat to achieve goals) and the “monitorability tax” (trading raw performance for transparency) we might all want to start paying.

The Hidden Cost of AI Agents No One Talks About (w/ Darin Patterson)

Make's VP Darin Patterson argues most companies fail with AI agents because they lack solid automation foundations. Before deploying autonomous agents, you need deterministic workflows to handle rules and logistics. The episode demystifies Make for those absolutely beginning to use AI, and previews Make's new AI capabilities to help you make, well, Make workflows codenamed “Maia.” (coming soon!).

Why IBM Wants AI to Be Boring (w/ David Cox)

IBM's David Cox argues AI should be treated like electricity (boring, reliable, ubiquitous) rather than a “friendly robot.” We discuss IBM's Granite 4.0 open-source models, which use hybrid architecture to drastically reduce memory and cost, and why we might not actually want AGI for most business applications.

And if you haven’t subscribed yet, please do! Click the image below to go to our channel and hit “subscribe” to get notified right when new videos go live.

We have a goal to hit 50K subscribers by the end of the year (if not 100K), and we’re only ~36K away! If you like learning about AI, and already watch some of our videos, do us a favor and click here to subscribe today.

Stay curious,

The Neuron Team

P.P.S: Love the newsletter, but don’t want to receive these podcast announcement emails? Don’t unsubscribe — adjust your preferences to opt out of them here instead.