Meta details 30B open-weight Muse Glimmer for one consumer GPU

#1 𝕏

Meta released the Muse Glimmer technical report, detailing its 30B open-weight model for agentic tasks that runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision. Meta says it leads comparably sized open models in tool use and multi-step planning.

#2 🐙 anthropics/claude-code

anthropics/claude-code - anthropics/claude-code is Anthropic’s agentic coding tool for using natural-language commands to execute routine tasks, explain code, and handle Git workflows in a terminal, IDE, or GitHub. On October 19, 2026, Anthropic released Claude Code v2.1.304, adding auto-mode setup, mod permission controls, a Managed Agents security check, and fixes; the repository also contains plugins with custom commands and agents, plus examples. Worth trying if you need a codebase-aware assistant with extensible workflows. Release notes

#3 𝕏

Cognition shared how it built Fusion, which was available to try in the Devin CLI at the time of the post. The post links to Cognition’s article and the Devin CLI.

#4 📝 Ampcode Chronicle

Assign It To Amp - Amp is now available as an agent in Linear: assigning it an issue or mentioning @Amp starts an Agent Session, where Amp responds to messages. Sessions also appear in the Amp sidebar and are fully viewable in Amp.

#5 𝕏

OpenAI shared model advances that strengthen software development and migration in low-resource languages including COBOL, Fortran, and Visual Basic, helping developers understand decades of business logic and modernize systems while preserving what works. It is seeking input on which languages and workflows to address next.

#6 🐙 pbakaus/impeccable

pbakaus/impeccable - **pbakaus/impeccable** provides design guidance for AI coding agents through 1 skill, 24 commands, live browser iteration, and 61 deterministic detector rules that require no LLM or API key. pbakaus released CLI v4.7.0 on October 19, 2026, adding local debugging-port support for desktop and webview apps, window targeting, and engine 0.1.16 typography checks. Worth trying if you need repeatable frontend design guidance and automated checks. Release notes

#7 𝕏

Guillermo Rauch said an HTML artifact builder powered by libfx was released as free, open-source software at fx0.vercel.dev, with plumbing designed for v0 to survive server crashes, flaky LLMs, and function limits. He also demonstrated an HTML floating container that he built in a few minutes.

#8 🐙 coreyhaines31/marketingskills

coreyhaines31/marketingskills - **coreyhaines31/marketingskills** is an MIT-licensed library of reusable marketing skills for AI agents, covering CRO, copywriting, SEO, analytics, and growth engineering. On October 19, 2026, coreyhaines31 released v2.13.0, adding Aaron Beashel’s conversion-tracking skill v1.0.0 with nine references, 26 detection snippets, importable GTM recipes, and 16 evals; the repository now has 52 skills. Worth trying if you need agent workflows for conversion tracking across major advertising and analytics platforms. Release notes

#9 🐙 heygen-com/hyperframes

heygen-com/hyperframes - heygen-com/hyperframes is an Apache-2.0 open-source framework that turns HTML, CSS, media, and seekable animations into deterministic MP4 videos, with a CLI and 21 published skills for AI coding agents. heygen-com released v0.8.181 on October 20, 2026, fixing stalled or timed-out long phone-video exports and parallelizing true-peak checks; at discovery, it had 68,005 stars and 5,945 weekly trending gains. Worth trying if you need agent-driven, HTML-based video rendering. Release notes

#10 🐙 mvschwarz/openrig

mvschwarz/openrig - mvschwarz/openrig lets builders define persistent Claude Code and Codex agent teams in YAML, with a CLI, reusable skills, docs, demos, and Docker files. mvschwarz released OpenRig v0.6.13 on 2026-10-20 at 05:39:24Z with 46 new changes, including 15 from 10 outside contributors. Worth trying if you need shared context, roles, and owned work across coding agents. Release notes

#11 🐙 MadsLorentzen/ai-job-search

MadsLorentzen/ai-job-search - MadsLorentzen/ai-job-search is an MIT-licensed local framework built on Claude Code for finding jobs, evaluating postings, tailoring CVs and cover letters, and preparing interviews, with reusable agents, skills, scrapers, templates, and tools. MadsLorentzen released v1.7.3 on October 19, 2026, tightening permissions and screening postings for hidden instructions after contributors closed 58 bugs in two weeks. Worth trying if you need a customizable workflow and can use Claude Code and Python 3.10+. Release notes

#12 𝕏

AI at Meta recapped training its Muse foundational models, including Muse Wonder, on Prometheus, its gigawatt-scale supercluster in New Albany, Ohio, with hundreds of thousands of GPUs. During Muse Wonder training, Prometheus achieved 85%+ goodput on pretraining, 90%+ on reinforcement learning, and ~47K GPU hours mean time between interruptions.

#13 𝕏

Jason Zhou said his team created an unnamed open-source tool that finds AI-cited sources, their authors, and their email addresses after four AI products cited the same ~40 pages—none belonging to his team. He said it costs 1¢ a prompt and mentioned partnering with @cloroai.

#14 𝕏

Philipp Schmid recapped Google Research’s Harness², a framework that improves the harness around a frozen model by testing 2 candidate edits, merging fixes, and saving verified lessons and reusable components. Reported results for Gemini 3.5 Flash include +8.4 on WorkBuddy-Bench and +7.5 on JobBench, improvement on all 3 benchmarks, and skills and scripts driving approximately 50% of edits—with dropping skills costing 5.3 points.

#15 𝕏

LlamaIndex 🦙 announced that it trained additional speculative decoders to reduce latency for OSS models on OpenDocRouter, with considerably faster inference expected for models such as LightOnIO’s LightOnOCR-3-1B. It measured tokens per second for single requests (best case) and 64 requests (saturated throughput), noting that TeleOCR sees less speedup at scale because it makes many small calls where decoding latency matters less.

#16 🐙 thedotmack/claude-mem

thedotmack/claude-mem - **thedotmack/claude-mem** is an Apache-2.0 persistent-memory system that captures agent tool use, creates semantic summaries, and restores relevant context in future sessions; the repository includes an installer, plugins, integrations, documentation, and an SDK/library. thedotmack released v14.0.1 on October 17, 2026, updating cards to retain the agent responsible for work when new updates arrive. Worth trying if you need continuity across AI agent sessions. Release notes

#17 🐙 Panniantong/Agent-Reach

Panniantong/Agent-Reach - **Panniantong/Agent-Reach** is an MIT-licensed command-line capability layer for giving AI agents reading and search access across services including YouTube, GitHub, Reddit, Twitter/X, Bilibili, and Xiaohongshu. The repository includes source code, configuration, documentation, scripts, and tests, with diagnostics and fallback backends; some platforms require login or configuration. Worth trying if your agent needs managed access to multiple internet sources.

#18 🐙 browserwing/browserwing

browserwing/browserwing - **browserwing/browserwing** is an MIT-licensed browser automation project that provides structured website data through an agent-focused CLI, MCP, and Skills support. It includes 78 built-in scripts across 10 categories, a visual recorder, and 26+ HTTP API endpoints, and requires Chrome or Chromium. Worth trying if you need reusable browser workflows for AI agents.

#19 🐙 morluto/rea

morluto/rea - morluto/rea released REA (rea-agents) 6.7.0 on October 18, 2026. This MIT-licensed MCP server and CLI lets agents locally inspect native binaries, JavaScript and Electron apps, .NET assemblies, and websites, returning evidence and limitations; the release also improves duplicate JSON/JSONC configuration handling. Worth trying if you need evidence-based reverse engineering across application layers. Release notes

#20 🐙 langfuse/langfuse

langfuse/langfuse - **langfuse/langfuse** is an open-source platform for tracing, evaluating, monitoring, and debugging AI applications, with prompt management, datasets, a playground, APIs, and cloud or self-hosted deployment. The 36,073-star repository’s owner, langfuse, released v4.64.0 on October 19, 2026, renaming Langfuse Assistant to Halo and adding transcript comparison, run costs, and API-key changes. Worth trying if you need one platform to inspect and improve LLM apps. Release notes