Watch on YouTube · subscribe for the daily brief

“A closed model can be retired on someone else’s schedule. Weights you’ve downloaded stay on yours.”

On Monday we said Meta had closed the door on open weights. On Thursday, Meta shipped the weights for an eight hundred sixteen billion parameter model.

It’s called Muse Spark Open. It’s a mixture-of-experts model that activates about forty-two billion parameters per token. It takes text, images, video and audio, holds a one million token context, and was trained for long agent work like coding. The weights are on Hugging Face in two versions, full precision and a smaller NVFP4 build. Meta describes it as the open version of the Muse Spark model it released in August, with extra safety training for standalone use.

To be precise about Monday’s claim: Meta’s frontier models, Muse Wonder and Wonder Full, are still closed. Spark Open sits a tier below them. But it’s a large, capable model you can download today, and that matters.

The surprise is where you can run it. Meta and Apple tuned it for Mac Studios. The NVFP4 build fits across two M5 Ultra machines with five hundred twelve gigabytes of memory each, or four with half that. On servers, it runs on four Blackwell Ultra GPUs or eight H200s. It’s also live on hosted providers including Fireworks, Groq, Together and OpenRouter.

Now the license. It’s the Muse Model License, not Apache. It lets you use, fine-tune, distill and redistribute the model. But products must display “Built with Muse Spark Open.” And hosting providers earning over twenty million dollars a year from serving it need a separate deal with Meta. Internal use and your own product features are exempt from that revenue clause. Read the text before you ship.

Meta wasn’t the only one. Nvidia released Nemotron 3.5 Super VL, an open model built to run enterprise agents nonstop, and squeezed a seventy-four billion parameter agent model into twenty-four gigabytes so it can run on a laptop. Black Forest Labs’ open FLUX 3 Image Klein runs on consumer GPUs with eight gigabytes of memory. And a startup called nFox says it will open-source its sandbox for running open models.

So the open option got bigger and smaller on the same day. Bigger at the top, with an eight hundred billion parameter model you can host yourself. Smaller at the edge, with agent models that fit on a laptop.

Here’s why that matters this particular week. On Wednesday, OpenAI told developers it will shut off GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini in its API on January fourteenth. That includes fine-tuned models built on them. OpenAI’s own policy promises at least six months’ notice for generally available models. This notice gives three. The policy allows shorter notice for safety or compliance reasons, but OpenAI hasn’t said either applies. And the suggested replacement, GPT-6 Luna, handles reasoning and tool calls differently enough that swapping the model ID may be the easy part.

That’s the trade in one week. A closed model can be retired on someone else’s schedule. Weights you’ve downloaded stay on yours. You pay for that with hardware, operations and a license to read. But nobody can send you a three-month notice.

So what do you do? If any production workflow depends on a model you don’t control, work out your migration window if the notice came tomorrow. Then pick one open model as a tested fallback for that workflow, and actually run your evals on it. A fallback you’ve never tested is only a hope.

Builder’s bench

• Anthropic is giving METR researchers employee-level access to look under the hood, and it made another acquihire aimed at confidential AI. Bank of America put four thousand people on fixing vulnerabilities found by Anthropic’s Mythos and other AI models. Ars reports the FTC probe into Anthropic and OpenAI went ahead after Trump dismissed AI fears as a hoax.
• Eight states sued OpenAI over harms to children, per CNBC. GPT-6 Astra solved forty-nine of fifty physics research problems at about two dollars forty each. And ChatGPT can now peek at your Chrome tabs.
• On the data side, CedarDB pitches one database for transactions and analytics, and Pinecone added an on-demand mode for bring-your-own-cloud. Databricks shared lessons from sizing Lakebase search, and Snowflake launched sovereign cloud and AI offerings in France. The New Stack argues agentic AI has a retrieval problem more than a model problem, and Ben Rogojan asks whether data teams are solving the right problems.
• Security: GhostAction credential theft hit nineteen hundred GitHub accounts and swept thirty thousand repositories. BleepingComputer asks whether AI pentesting matches a human. The New Stack says the permission you deny today is your safety net tomorrow.
• Hardware: TSMC posted record profit for a sixth straight quarter. Cerebras is pitching inference speed for agent workflows, and Windows is being tuned for Nvidia’s RTX Spark laptops. Jensen Huang called for US-China dialogue on potentially dangerous AI.
• Vercel added always-on workers for backends. LangChain answered the most common questions about agent evals. And InfoQ has a talk on what twenty-five trillion tokens reveal about how developers adopt agents.
• Also: Google wants Gemini 4 Argon to be the only model robots need, Amazon’s Bee wearable can now help you shop, and YouTube Music will label AI-generated songs.

The through-line: the open option got bigger and smaller this week, right as a closed model got a short notice, so keep a tested fallback you control.

One thing to do today: list every production workflow pinned to a specific hosted model. For each one, write down the open model you’d move to and the date you last tested it. Every blank date goes on your to-do list.

Sources covered

1 story · 77 sources (64 news, 10 company, 3 voices)

1. 8 insights from LogicMonitor’s Elevate event: AI-driven observability moves toward autonomous IT
   📰SiliconANGLE↗📰CNBC↗🏢Snowflake↗🏢Snowflake↗🎙️Simon Willison↗📰TechCrunch↗📰The Verge↗📰SiliconANGLE↗🎙️Ben Rogojan↗📰SiliconANGLE↗📰The New Stack↗📰TechCrunch↗📰The Verge↗📰TechCrunch↗📰BleepingComputer↗📰The Hackers News↗📰TechCrunch↗📰TechCrunch↗📰The New Stack↗📰The Decoder↗📰The Verge↗📰TechCrunch↗📰The New Stack↗📰Hacker News↗📰SiliconANGLE↗📰SiliconANGLE↗🎙️Simon Willison↗📰The New Stack↗📰The Decoder↗📰The Decoder↗📰The Verge↗📰The Decoder↗📰The New Stack↗🏢Pinecone↗📰SiliconANGLE↗📰Ars Technica↗📰SiliconANGLE↗📰The Decoder↗📰The Decoder↗📰Ars Technica↗📰TechCrunch↗📰The New Stack↗📰The New Stack↗📰The Verge↗📰TechCrunch↗🏢ClickHouse↗📰The Verge↗📰BleepingComputer↗🏢NVIDIA↗📰TechCrunch↗📰The Verge↗📰The New Stack↗📰The Hackers News↗📰CNBC↗📰LWN.net↗📰SiliconANGLE↗📰The Verge↗📰The New Stack↗🏢Anyscale↗🏢Starburst↗📰TechCrunch↗📰The Verge↗📰The New Stack↗📰The Verge↗📰InfoQ↗📰CNBC↗📰Lenta.RU↗📰The Hackers News↗📰The Decoder↗🏢NVIDIA↗📰TorrentFreak↗📰TechCrunch↗📰Lenta.RU↗🏢LangChain↗📰CNBC↗📰TechCrunch↗🏢Databricks↗