Meta Opens Muse Spark Weights as Evaluator Access Becomes Norm — Oct 16

Meta opens Muse Spark weights

Show Notes

Meta has released Muse Spark Open, an 816B open-weight model that leads US open weights but trails three Chinese models on Artificial Analysis. Anthropic gave METR employee-like access to its systems and set three evaluator tiers, while the FTC reportedly sent 50-page demands to Anthropic and OpenAI over safety claims and incidents. A Goodfire essay argues neural networks are already clean, and Google and Amazon pushed agent checkout and sub-$100 AI wearables.

What is The Harness?

A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.

Good morning, it's Friday, October sixteenth.

In today's briefing we see Meta releasing the weights of Muse Spark Open, Anthropic formalizing outside evaluator access with METR, and reported FTC demands to Anthropic and OpenAI over safety claims.

First up - Today in the big model news;

Anthropic

Anthropic has given six METR employees company hardware, office space and internal tools, roughly on par with its own risk-assessment teams. Customer data and model architectures are excluded. METR can publish its analysis without a veto, though Anthropic can require redactions in narrow categories, and METR may say whether it thinks they were justified. Two lower tiers cover narrower evaluators, one needing approval to publish and one not. SecureBio is already reviewing the chem and bio chapter of the August Risk Report. Anthropic says it signed the White House Accord on Super Intelligence, and that every frontier company should. This is the compliance plumbing for government-pressed access rules. Until OpenAI and Google publish comparable tiers, enterprise security reviews can reasonably ask which evaluators hold what access to a model supplier.

Meta

There is a continuing question over whether frontier models are becoming commodities, and Meta's latest release bears on it. Meta has released Muse Spark Open, an open-weight model, two days after its closed Muse Wonder models.

It has eight hundred sixteen billion parameters in a mixture-of-experts design, with forty-two billion active and a one-million-token context window. The license is limited-commercial.

Artificial Analysis scores it thirty-nine. That puts it behind MiMo-V2.6-Pro, GLM-5.3 and Kimi K3, and ahead of Mistral Large 4. It also calls the model particularly expensive for an open model. So Meta's claim to the most capable US open-weight model holds only with the US qualifier.

The license requires attribution, and per secondary reports, a separate agreement above twenty million dollars in third-party API revenue. That clause targets rival API businesses, not users.

Seven inference hosts and three hyperscaler catalogs already serve the weights. Meta is competing on provenance for buyers barred from Chinese weights, and Mistral is the most exposed. Meta's claim that two Mac Studios can run it is vendor-reported, and TechCrunch says it needs at least one terabyte of RAM. Compare hosted per-token prices against the Artificial Analysis cost figures before committing.

On the regulatory front;

Gizmodo, citing Axios and one unnamed source, reports the FTC sent civil investigative demands to OpenAI and Anthropic on October fourteenth, each over fifty pages. They reportedly seek documents on security incidents, safety measures, marketing materials and privacy policies. The agency reportedly plans to order executives to testify and to issue demands to third-party AI organizations. The details are unverified, and neither company nor the FTC had responded. On September twenty-fifth, FTC Chair Ferguson said developers instructing agents in cyber tests should be liable for harm. Marketing language about safety is now potential evidence. Review how a product describes its safeguards with the same care as the safeguards, and keep incident logs retrievable.

In other news…

Goodfire's Nick Cammarata argues that apparent polysemanticity mostly comes from fixable training problems, not superposition. The essay itself labels its central claim, that networks are fully interpretable, highly speculative. A preliminary audit of Qwen3-32B finds about ninety percent of neuron poles have one main job, though the author says the figures depend heavily on thresholds. Squeezing one InceptionV1 layer from four hundred eighty neurons to ten kept about fifty-eight percent top-one ImageNet accuracy with still-interpretable neurons, which he reads as evidence against superposition. It is an informal, non-peer-reviewed essay. If it holds, direct neuron-level monitoring could undercut the probe-based approach behind Goodfire's own vendor-reported agent monitors. Treat it as a hypothesis, not a roadmap input.

Quick hits from the consumer side;

Google's Universal Cart is live in the US across Search and the Gemini app, saving products from multiple retailers into one cart and checking out through Google's Universal Commerce Protocol.

Amazon's Bee launched Bee Frontier at seventy-four dollars and ninety-nine cents and Bee Eclipse at ninety-nine dollars and ninety-nine cents, with the useful tier a Bee Pro subscription at eleven dollars and ninety-nine cents a month that bundles Alexa+ and Amazon.com integrations.

That's the briefing. Have a great day, and don't forget to subscribe.