Meta ships Muse Glimmer, a 30B open-weight model built to run and act on one GPU

Good morning 👋 Meta just open-sourced a 30B agentic model under Apache 2.0, and the "runs on your laptop" headlines are already lying to you about which build they mean.

In today's issue:

• 🦾 Meta ships Muse Glimmer, a 30B open-weight model built to run and act on one GPU
• 🧠 NVIDIA's Cosmos 3 opens up physical-AI world models in three sizes
• 🔬 A context eviction trick, a fix for self-improving skills, and distillation on one GPU
• 🛠️ Put a session leash on the agent you just gave more autonomy
• 🔗 A security robot's speed-bump problem, a retailer's eighth good quarter, and a cash-flow test to watch

🔭 THE ONE THING

🦾 Meta ships Muse Glimmer, a 30B open-weight model built to run and act on one GPU

Meta's Superintelligence Labs dropped Muse Glimmer today: Apache 2.0, 30B open-weight params, tool-calling, object detection, and image/video input baked in. The "runs on your laptop" framing from Meta's own blog and Alexandr Wang's X post is about the quantized build, not this one, so read the specs before you provision. Full fine-tuning still needs serious multi-GPU infrastructure, which tells you who this release is actually for: not hobbyists, but any team that wants an agentic, tool-using vision-language model they can own outright instead of renting through an API. Apache 2.0 on a frontier-adjacent open model from Meta, two years after everyone assumed Llama's open era had peaked, is the real headline here. Watch what ships on top of it in the next month, that's where you'll know if this is a genuine platform or a one-week Hacker News spike.

🧠 MODELS & RELEASES

• 🤖 NVIDIA shipped Cosmos 3, its physical-AI world model family, in three open-weight sizes: Edge (4B), Nano (16B), Super (64B). It's already topping benchmarks for world generation and robot-policy tasks, and it ships under the OpenMDW 1.1 license, so Agile Robots, Black Forest Labs, Generalist, LTX, Runway, and Skild AI can post-train it on their own robot data instead of waiting on NVIDIA. NVIDIA

🔬 RESEARCH HIGHLIGHTS

• Your coding agent is hoarding context it'll never need again, and now there's a way to tell which. Blast Radius adds a memory layer that predicts how far an incoming prompt actually reaches through your code and chat history, then evicts what falls outside that radius, archiving it verbatim instead of deleting it, so it's recoverable if you need it back. Tested across seven OpenAI models, it posted the lowest overflow rate of any policy compared. Of 450 archived items, 378 turned out to be recurring dead weight that got recalled exactly zero times. paper

• Most "self-improving" agent skills can't tell the difference between editing a skill and killing it. SkillProx splits the two: a forward pass diagnoses failures and edits with rollback, a backward pass decides what to consolidate, demote, or cut, borrowing the proximal-gradient idea from optimization and pointing it at text instead of weights. It beats the strongest gradient-based skill-refinement baseline on accuracy, and the gain holds across model backbones and both in-distribution and out-of-distribution tasks. Skill libraries compound, and sloppy deletion is exactly the kind of error that piles up quietly. paper

• Distillation's bottleneck was never compute. It was memory. Multiverse Computing caches a teacher model's top-100 logits once instead of re-running it on every batch, then chunks the KL-loss computation instead of materializing the full vocabulary matrix. That memory savings let them distill a GPT-OSS 20B model on one GPU node instead of four, at nearly 5x the throughput. The resulting 3.2B student stayed within about nine points of its 8B teacher on MMLU. Distillation just got cheap enough for teams without a GPU farm to bother. blog

🛠️ TRY THIS

Put a session leash on the agent you just gave more autonomy

Muse Glimmer runs locally and takes agentic actions on your own hardware. That's freedom you didn't have to ask anyone for, and freedom is exactly what needs a leash. AWS just published its version of that leash for Bedrock AgentCore. The pattern travels even if you're nowhere near AWS.

1. List every tool or action your agent can call. Tag each one: read-only, reversible, or irreversible (spending money, sending an email, deleting a file).

2. For irreversible actions, gate on session history, not identity alone. "Has this agent already gotten a human sign-off in this session?" beats "does this agent have permission?"

3. Cap request and token volume per session so a bad loop can't quietly burn your API budget or hammer a downstream tool.

4. Try to trick it. Run a session where you skip a required step and confirm the agent refuses the irreversible action.

Prompt: List every tool my agent can call. For each, classify as read-only, reversible, or irreversible, then draft a session-history rule that blocks the irreversible ones unless a prior step in the same session already confirms authorization.

Worth a look

• AWS Bedrock AgentCore ships temporal policies that gate actions on session history instead of a one-time permission check. Securing AI agents with temporal policies
• AWS Bedrock AgentCore gateway now takes per-user and per-target rate limits on requests, tokens, and connections. Configure rate limits for AI traffic

🔗 QUICK LINKS

• AITX filed an 8-K just to explain why its ROAMEO security robot kept bonking into a speed bump on camera. CEO Steve Reinharz calls the repeated nudges "a deliberately conservative control strategy," not a malfunction. 8-K
• Citi Trends pre-announced Q2 sales up 10.9% to $211.6 million, its eighth straight quarter of comp growth. No AI angle, just a discount retailer having a good quarter. 8-K
• Boost Run set August 14 for its Q2 print, the real test of whether $940 million in contracted revenue is turning into cash. 8-K

That's the signal worth your time today.

Pradeep Perugu