🦙 Meta puts Muse Spark Open in builders’ hands

Meta released Muse Spark Open, an open-weight multimodal model built for coding and agent work. The meaningful shift is access to the weights, with desktop execution already demonstrated on a high-memory Mac. Read the original post on X →

• The model has 816B total parameters, 42B active parameters, and a 1M-token context window.
• Meta lists native support for text, images, video, and audio.
• A practitioner demonstrated llama.cpp execution on an M3 Ultra with 512GB of memory. That is a substantial desktop, not an ordinary laptop.
• OpenRouter lists hosted access at $1.25 per million input tokens and $4.25 per million output tokens. Open weights should not be read as unrestricted commercial permission.

What it means: You can run Meta’s new agent-focused model outside a hosted API, but hardware requirements and license terms still need checking before deployment.

📦 Perplexity releases the sandbox behind Computer

Perplexity has open-sourced SPACE under Apache 2.0, making the platform that runs every Perplexity Computer task available to outside developers. This is production infrastructure becoming reusable, not just another sandbox demo. Read the original post on X →

• Each task gets its own Firecracker microVM, with credentials and outbound traffic controlled from outside that virtual machine.
• Agents can run commands and install packages with root access inside the sandbox; a separate gateway enforces outbound connection policy.
• Sandboxes can pause, fork, or suspend to memory and disk snapshots, then resume their work.
• Perplexity reports that nine model configurations made 108 escape attempts without reaching the host. That test is evidence, not a containment guarantee.

What it means: You can reuse a production agent-isolation platform with external credential and network controls instead of building those boundaries entirely inside an agent’s environment.

🔎 METR gets employee-like access inside Anthropic

Anthropic has begun giving six METR employees deep internal access to assess its risks and safety practices. Unlike the broader review proposal covered yesterday, this arrangement is already operating and includes publication rights. Read the original post on X →

• METR’s team has company hardware, office space, and access roughly comparable to Anthropic’s own risk-assessment teams.
• Exceptions include customer data, attorney-client privileged information, and model-architecture information.
• METR can publish assessments of risk levels, incidents, and practices without Anthropic vetoing publication.
• Anthropic can require narrowly defined redactions, but METR can publicly describe those redactions and whether it considers them justified. The arrangement is voluntary, not regulatory supervision.

What it means: Frontier-lab safety claims now have a concrete path to outside examination, though the value of that scrutiny will depend on the assessments METR publishes.

💾 Vercel lets agents carry files between sandboxes

Vercel Drives is generally available for Vercel Sandbox on all plans. It keeps an agent’s files separate from the environment running its code, so changing runtimes or compute no longer requires transferring the workspace. Read the original post on X →

• Create a drive once and mount it into a sandbox. Files on that drive persist when a new sandbox mounts it.
• Other agents can mount read-only copies of a prepared dataset and save their results separately.
• Writable forks let agents test different approaches independently. Forks do not merge back into the original drive.
• Default capacity is 1 TiB, or 1 GiB on Hobby, with up to 16 TiB available.

What it means: Keep durable agent workspaces separate from disposable compute, then reuse or fork their files when the next task needs a different environment or parallel experiments.

How was this issue?