← ALL THOUGHTS

OCT 16, 2026 · AI BRIEF

Open Weights Are Back, With a Catch

Jeremy's assistant avatar

From Jerm's desk, via his assistant

✓ APPROVED BY JEREMY

The short version: Meta released a big open-weight model anyone can download, and the catch is the hardware it takes to run it. RTX Spark laptops built for local AI go on sale today, and agents are moving into cars and office processes with the limits stated out loud.

Jeremy's illustrated character in a navy blazer and white tee at his desk looking at two desktop computers side by side showing gold network diagrams, a gold spark rising above them and an open gold box beside his papers.
A big open model you can download today, running across two desktops on a desk.

TL;DR

  • Meta released Muse Spark Open on Oct 15, an open-weight model with 816 billion total parameters and a 1 million token context window, built for coding and agent work. You can download it free, but running it locally at full size takes about a terabyte of memory, like two high-end Mac Studios linked together.
  • Local AI hardware has a real day today. The first NVIDIA RTX Spark laptops go on sale Oct 16, starting at $2,599 in the US, with up to 128 GB of unified memory for running models on the device.
  • Agents are moving into new places with clear limits. Rivian says its cars will one day run third-party agents, but steering and braking stay out of agent hands, and Microsoft is letting Copilot Studio agents start work on their own in its Frontier program, with their own accounts and scoped permissions.

Top stories

Meta puts open weights back on the table with Muse Spark Open

TECHCRUNCH AND REUTERS, OCT 15

Meta released the weights for Muse Spark Open, a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active per token, and a 1 million token context window. Meta calls it the most capable US open-weight model to date. It is built for long coding tasks, multi-step reasoning, and agent work. The weights are free to download from places like Hugging Face, with a license that allows limited commercial use, and hosted APIs are live through providers including Together AI, Fireworks AI, Groq, and OpenRouter. Meta worked with Apple so the model can run locally across two Mac Studios with M5 Ultra chips and 512 GB of unified memory each, using Apple's MLX framework.

Why it matters: Open weights give builders control. You can inspect, fine-tune, and run a model without sending everything to someone else's cloud. The honest catch is in the same story. Independent comparison coverage citing the Artificial Analysis index puts Chinese open models like GLM 5.3 ahead of Muse Spark Open overall, and Meta's own strongest model, Muse Wonder, stays closed. So this is a real step for US open weights, not a finish line. And local at this size still means serious hardware, not a normal laptop.

Read the source

RTX Spark laptops go on sale today

TECHTIMES, OCT 16

The first laptops built on NVIDIA's RTX Spark chip go on sale Oct 16 from makers including ASUS, Dell, HP, Lenovo, MSI, and Microsoft, whose entry is the Surface Laptop Ultra. The chip pairs a Grace CPU with a Blackwell GPU and up to 128 GB of unified memory. US entry pricing starts at $2,599 for a 24 GB configuration, and reviews landing today praise the GPU while flagging weaker CPU performance on battery as a qualifier.

Why it matters: Put this next to the Muse story and the pattern is clear. The industry is building hardware specifically so capable models can run where you are, not only in a data center. Whether that is worth $2,599 and up is a real question, and the first reviews are the place to start, not the launch pitch.

Read the source

Rivian wants other companies' agents in its cars, with limits

TECHCRUNCH, OCT 15

At TechCrunch Disrupt, Rivian CEO RJ Scaringe said future Rivian vehicles will be able to run AI agents from other companies, naming Anthropic as an example. No timeline and no confirmed partners yet, and Rivian pointed to its AI and Autonomy Day in December for details. Rivian already has its own voice assistant and an in-house system that routes simple requests to a local model and harder ones to the cloud. Reporting on the talk notes vehicle controls like braking and steering are not being handed to an agent.

Why it matters: The car is becoming another place an agent can act for you, with your calendar and context along for the ride. The useful part of this announcement is the boundary. Help with planning and information is one thing. Control of the car itself is another, and Rivian is saying that line stays.

Read the source

Google launches Universal Cart across Search and Gemini

UNITE.AI, OCT 15

Google rolled out Universal Cart in the US across Google Search and the Gemini app. It saves products from multiple retailers into one cart tied to your Google account, watches for price drops and restocks using Google's Shopping Graph, surfaces promo codes and card perks through Wallet, and lets you check out with Google Pay at supported stores or move items to the merchant's own cart. The merchant stays the seller of record.

Why it matters: This is the calmer sibling of the agent shopping story from earlier this week. The agent does not go buy on its own. The cart does the watching and organizing, and the person still decides and pays. That split, automation for tracking and a human for the purchase, is probably how most people will want this to work for a while.

Read the source

Microsoft lets Copilot Studio agents start work without a prompt

FRONTIER BLOG, OCT 16

Microsoft announced new Copilot Studio capabilities in its Frontier program that let agents take on shared business processes on a trigger, like an arriving email or a schedule, instead of waiting for someone to open a chat and prompt them. The agents get their own accounts and scoped permissions in Microsoft 365, can be reached in Teams and Outlook, and carry memory of the process from one cycle to the next.

Why it matters: This is the work version of the same move Rivian described for cars. Agents stop being only chat windows and start being assigned workers inside a process. That only works if the account, the permissions, and the audit trail are clear, which is exactly the part to read closely before turning it on for anything real.

Read the source

Signals from leaders

THEIR OWN POSTS AND STATEMENTS, LABELED AS SIGNALS

Mark Zuckerberg (Meta), X post, Oct 15: Zuckerberg announced the Muse Spark Open weights release and credited the work with Apple and its CEO John Ternus for how well it runs on new Mac Studios. Apple's Ternus replied publicly the same day. Labeled as the leaders' own posts, reported by MacRumors. Read the coverage

Alexandr Wang (Meta chief AI officer), X post, Oct 15: Wang called Muse Spark Open the most capable US open-weight model to date. Other coverage notes he also acknowledged it is not Meta's most powerful model, since Muse Wonder stays closed. Treat the ranking claim as Meta's claim, checked against independent indexes in the top story above. Read the coverage

Andrew Ng, interview reported Oct 15: Ng pushed back on extinction level safety arguments in an Economist podcast interview, arguing there are no meaningful existential risks from AI and that some safety claims are used to push regulation that hurts open competition. That is a direct counterpoint to the pacing argument other leaders backed last month, and worth hearing in his own lane. Labeled as interview coverage, not a new issue of The Batch, and no new Batch letter was found in the window. Read the coverage

Satya Nadella (Microsoft): No new post from Nadella in the last 24 to 36 hours. An Oct 15 NBC News report, covered in follow-up writing, pairs his push to make Windows the safest place for agents with his Oct 10 warning not to trust models uncritically. That essay was already covered here, so this is context, not a new signal. Read the coverage

Jack Clark (Anthropic, Import AI): No new Import AI issue since 476 (Oct 12). Clark is in Bangkok for the IMF and World Bank annual meetings, where on Oct 16 the IMF chief said AI risks could exceed the opportunities without safeguards and an ethical framework, in a panel discussion that included Clark. He is also scheduled to give the Per Jacobsson lecture on Oct 17. Read the coverage

All-In Podcast: The newest listing is an Oct 15 conversation with Billy Evans, which is not an AI leader episode. The last AI heavy episode remains Oct 10, already noted here.

No verifiable new items in the window for Andrej Karpathy, Boris Cherny (last verified posts are Oct 6 to 9), Allie K. Miller, Sam Altman as a direct post, Dario Amodei as a direct post, Elon Musk on AI specifically, or Jensen Huang. OpenAI and Anthropic both appear today through the Simons Foundation partnership in Go Deeper.

New models, tools, and hardware

  • Muse Spark Open, practical routes: Do not start with the 1.6 TB full precision download. The compressed version is still hundreds of gigabytes. For most builders the hosted APIs are the real on-ramp today, with the two Mac Studio setup as a proof of what local can do at the high end. Read the launch coverage
  • Amazon Bee Frontier and Eclipse (Reuters, Oct 15): Amazon's Bee launched two personal AI recorders. The Frontier is a wrist or clip device at $74.99 with up to seven days of battery. The Eclipse is a credit card sized device that attaches magnetically to a phone, from $99.99. Both turn conversations into summaries and to-dos. Amazon says raw audio is processed in real time and then deleted. The Verge's companion coverage this week is the right caution here: devices like this ask the people around you for data too, not just you, and consent is the hard part. Read the coverage
  • Amp pushes agents beyond code (OpenAI News, Oct 14): OpenAI's case study shows Amp using its models for work outside straight coding, another sign agent platforms are widening what a single session is expected to handle. Read the case study
  • NVIDIA powered city agents in Europe (NVIDIA Blog, Oct 14): Cities and partners in Europe are using NVIDIA accelerated computing for physical AI agents and digital twins for urban operations, from road repair timing to energy use. Less flashy than a laptop launch, and closer to how AI shows up in daily public services. Read the post
  • Microsoft lane check: Checked per the standing rule. The real new item is the Copilot Studio Frontier delegation update in Top Stories. The Maia 300 chip story is from Oct 12 and was already covered in this series, so it is not repeated as new. Muse Spark Open weights are also listed on Microsoft Foundry Models, noted in the Muse coverage above. Coverage parity held: Microsoft does not lead today's brief.

For builders

Today's arXiv papers all point the same way. The model is not always the problem. The harness around it is.

One paper proposes adapting the runtime harness, the layer that handles observations, tools, feedback, and control, instead of retraining the model, because many agent failures in rule based work come from a mismatch at that interface. Another, on agent memory, argues that finding memories that look similar is not enough. What matters is which memories actually help the next task, and how they are assembled and presented.

Plain version: before you blame a model or pay for a bigger one, check the setup around it. What can it see, what tools can it call, what memory does it get, and how do you know the work is done. That is builder work, and it is cheaper than a model swap. Harness paper · Memory paper

Career signal

Open weights are becoming a hiring skill again, but not in the hobbyist sense. Teams need people who can do three plain things well: evaluate an open model honestly against the vendor claim, decide what should run locally and what should stay in the cloud based on cost and privacy, and set the permissions for agents that now work in cars, in Windows, and in business processes. A portfolio project that shows an eval table, a cost comparison, and a clear permission boundary will read as serious work this month.

Go deeper

One project idea

Run an honest open model test. Pick one small coding or agent task you already do with a closed model. Run the same task through Muse Spark Open on a hosted API, not a local download, and write down four things: did it finish, how many steps did it take, what did it cost, and where did it need help. Then do the same task once more after you improve only the harness, meaning clearer instructions, better tool setup, or better memory. Publish the before and after table. That one page teaches evaluation, cost control, and agent design at the same time.

Closing

Yesterday was money and scrutiny. Today feels more practical. A big model anyone can download, laptops built to run models on your desk, agents invited into cars and office processes with the boundaries stated out loud. The question is shifting from can the model do it to where should it run, who gave it permission, and how will we check it. I am still learning that part in public, and I will keep sharing what I find.

If you want to talk about any of this, say so on Threads.

Sources for readers

About this brief: This brief was researched and drafted by my personal AI assistant and approved by me before publishing. Leaders' posts are labeled as signals and linked to the original coverage. Product details change fast; check primary sources before you build on them.

Get the week's briefs in one email: The Briefcase →