Meta releases 30B Muse Glimmer for always-on local agents

#1 𝕏

AI at Meta released Muse Glimmer, an open-weight, 30B-parameter model optimized for local, always-on agent workflows on consumer hardware, including Macs and PCs with performant GPUs. Its weights are available under the Apache 2.0 license.

Also covered by: @Alexandr Wang, @AI at Meta, @Alexandr Wang, @NVIDIA AI

#2 📝 OpenAI News

Expanding Daybreak as the Cyber Defense Window Narrows - OpenAI is expanding Daybreak into two tiers—Daybreak Blue, which provides trusted defenders access to GPT‑5.6 Sol with tailored safeguards removed to support vulnerability discovery, secure code review, malware analysis, incident response, and patch validation; and Daybreak Red, which provides purpose‑trained cybersecurity models for authorized vulnerability research, exploit validation, and security testing, including the new GPT‑5.6‑Cyber. OpenAI reports GPT‑5.6‑Cyber completes 95.0% of advanced cybersecurity requests on its internal Advanced Cybersecurity Completion Rate (versus 1.5% for GPT‑5.6 Sol and 2.0% for Sol with Daybreak Blue, and 57.3% for GPT‑5.5‑Cyber), outperforms earlier models on ExploitGym and a zero‑day benchmark, but produces shorter vulnerability reports and is less token‑efficient than GPT‑5.6 Sol on ExploitBench under a 300‑turn limit (the gap narrows at 600 turns).

Also covered by: @OpenAI, @OpenAI, @Sam Altman

#3 𝕏

Qwen demonstrated QwenLM’s Qwen-MM-Plugins, which makes agent harnesses multimodal-native for reading images, videos, and documents, editing videos, and working with 3D/CAD. The post links to the QwenLM GitHub repository to see it in action.

#4 📝 OpenAI News

Premium seats are coming to ChatGPT Business - OpenAI is adding Premium seats to ChatGPT Business that provide 5x the usage of Standard seats, remove the five-hour usage limit, and use predictable weekly usage resets; Premium seats cost $125/user/month ($100 billed annually) while Standard stays $25/user/month ($20 annually), and teams can mix and reassign seat types in one workspace. For a limited time the first 10,000 eligible workspace owners who join the waitlist can get $100 in workspace credits (2,500 credits) per Premium seat up to $500 for five seats and may receive early access; the promotion ends August 20.

#5 𝕏

Anthropic shared that an unreleased research version of Claude did not solve the Riemann hypothesis but increased the lower bound for the fraction of Riemann zeta function zeros satisfying it from 41.6% to 67.2%.

#6 𝕏

Claude Sonnet 5’s introductory pricing—$2 per million input tokens and $10 per million output tokens—is being made permanent and will remain unchanged after the previously stated August 31 end date.

#7 𝕏

Harrison Chase shared a tutorial showing how to use Stagehand v4—an SDK that lets agents browse the web from Browserbase—with Managed Deep Agents to create a production-ready web browsing agent.

#8 𝕏

Peter Yang recapped five takeaways from @thenanyu and @delashum at Linear for building production agents end to end: map the real workflow, equip agents to retrieve context, start with one frequent job, use the strongest model until the workflow works, and turn real failures into evals or product tasks. Linear’s first production workflow turned sales notes and Slack discussions into issues; the team launched it quietly and used observed behavior to guide subsequent workflows. Linear also created two feedback loops for poor agent behavior and missing-tool gaps.

Also covered by: @Peter Yang

#9 ▶️

Claude Code for normal people: skills, voice mode, and how to collaborate with AI

How I AI Podcast

#10 𝕏

Guillermo Rauch announced that Vercel Sandbox isolates ① compute with microVMs and ② network activity, characterizing OpenAI’s escape as occurring on the network path to Artifactory. Vercel’s Sandbox egress firewall is now free, enabling builders to further constrain agents’ network activity.

#11 𝕏

Jason Zhou introduced loopany.ai’s open-source loop library, featuring loops described as delivering real-world results and prompts users can copy to their agents.

#12 𝕏

Santiago shared that the specialized TwiL-LM family is now available on HuggingFace for Transformers and llama.cpp under a non-commercial license. The post says the 3B model beats OpenAI’s 120B gpt-oss on 4/5 formal-reasoning benchmarks while being 40x smaller and 2.6x faster in webAI throughput tests, while the 1.7B model is 1.06 GB and runs locally at ~367 tokens/sec.

#13 in

Guillermo Rauch recapped Vercel’s cloud-bill prevention functionality, including soft and hard caps, anomaly alerts, Functions recursion protection, agent-queryable billing APIs, and always-on L3/L4/L7 DDoS mitigation on all plans. Vercel’s streaming data infrastructure analyzes ingested data in real time to detect anomalies and threats and dispatch alerts.

#14 in

🥞 Carl Vellotti shared a five-level framework for agent memory, recommending builders stop at level 3—files plus a knowledge layer—and use CLAUDE.md as a table of contents rather than a filing cabinet. He also shared instructions for adding the fullstackpm CLI’s /docs-audit and /wiki-setup skills to Claude Code, Codex, or Cursor workspaces.

#15 ▶️

Cloudflare will make 1000+ AI millionaires

Greg Isenberg

#16 𝕏

Santiago shared a Hermes integration that will let developers embed Hermes into any application. Hermes could stream answers and reasoning live, render UI within the app, and act on each page.

#17 𝕏

Mustafa Suleyman announced that MAI-Image-2.6 is ranked the world’s #2 text-to-image model, ahead of Nano Banana, Meta, and Grok. It is available to try on Arena.

#18 𝕏

Thariq highlights two key skills for working with AI: allocating compute to problems worth pursuing and using AI as a thought partner. He says @__alpoge__ had to deeply examine and understand the proof to determine it was real.

#19 𝕏

Aravind Srinivas announced that a US-hosted K3 is available on the Perplexity Agent API.

#20 𝕏

Boris Cherny commented that he almost never hits rate limits on his personal account’s 20x, asking whether loops or routines might be involved. He suggested running `claude -p /usage | pbcopy` and sharing the output to debug, while requesting reproduction steps for the crashes.