Meta ships Muse Spark Open, an 816B MoE that runs on two Mac Studios

verifiedDeveloperCompute

Friday, October 16, 2026

Confidence

High · — primary release + multi-source corroboration; capability claim unverified

Evidence

Meta release + PyTorch blog + independent coverage

Meta released weights for Muse Spark Open — an 816B total / 42B active MoE with a 1M-token context — calling it the most capable US open-weight model, built for coding and agentic work.

The details

Runs on two Apple M5 Ultra Mac Studios via MLX, or four Blackwell Ultra GPUs.
Independent benchmarks pending; Meta's table shows mid-pack coding scores behind Kimi K3 and GLM 5.3.

What it means

A frontier-class coding model fitting on $20K of desk hardware changes data-residency more than the leaderboard. "Weights can't leave our VPC" no longer justifies a hosted-API deal — and Meta's "most capable US open-weight" framing concedes the open-weight race to China.

Meta released weights for Muse Spark Open — an 816B total / 42B active MoE with a 1M-token context — calling it the most capable US open-weight model, built for coding and agentic work.

Runs on two Apple M5 Ultra Mac Studios via MLX, or four Blackwell Ultra GPUs.
Independent benchmarks pending; Meta's table shows mid-pack coding scores behind Kimi K3 and GLM 5.3.

Sources

Apply this today

The hands-on layer the brief points to: a workflow you can run, a tool to test, and today’s 60-second video.

Workflow of the day

Audit Your AI Vendor's EU AI Act Exposure in 15 Minutes

Consulting & Enterprise · tested · saves 3 hours per vendor

Full walkthrough →

Tool watch

Anthropic Claude Developer Platform — Dynamic Workflows for Managed Agents

Claude's Dynamic Workflows: when the agent writes its own orchestrator

Read the breakdown →

In this issue

← Back to brief

Get the next one in your inbox

One daily brief. Every story gets a hype verdict.

No spam. Unsubscribe anytime.

On this page

The details
What it means

In this issue

← Back to brief

Stat of the Day

AI technology categories in Beijing's draft export catalogue

From large AI models and intelligent agents to chip design methodologies and computing cluster architectures — the scope touches nearly every layer of the modern AI stack. Source.

Today’s Take

The agent-containment story stopped being a vendor PR cycle this week and became three different kinds of policy artifact: a national CERT disclosure, a bilateral export regime, and a bank-industry protocol. Each is a different jurisdiction answering the same question — who is accountable when an autonomous system exceeds its brief — and each produces a different answer. The common thread is that "the vendor will tell us" is no longer a defensible control. Watch whether the EU AI Office or UK FCA publishes something analogous to CERT-In's note within 30 days; if they do, agent-incident disclosure becomes a global supervisory expectation before year-end. Considered and passed: the GE Appliances / Tata Steel 800-agent deployment counts (single thin-tier source, vendor-supplied metrics) and the UK FCA AI-supervision plan (announced intent, no mandate or timeline).

— Agentic desk

Never miss a signal

Join professionals getting the daily AI brief every weekday morning.

No spam · Unsubscribe anytime

No sponsored verdicts
Sources linked
Corrections public