Subscribe, if this was forwarded to you

Hey there,

This week brings a mix of AI policy shifts, open-source progress, and practical tools for builders. We cover the US export controls that forced Anthropic to suspend access to Fable 5 and Mythos 5, a new tool that automatically identifies the best local LLMs for your hardware, Microsoft's free machine learning curriculum, and a growing library of specialized AI agents for more reliable workflows.

Whether you're building AI applications, running models locally, or keeping up with the latest developments in the ecosystem, there's plenty to explore in this edition.

Enjoy the read!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Microsoft’s Machine Learning for Beginners course

🛠️ Tool Spotlight: whichllm auto-benchmarks local LLMs efficiently

📰 Top News: US export controls halt Anthropic Fable 5

💡 Tip: Boost results with Agency Agents prompt library

Let's get started!

Tutorial of the week

Master Machine Learning in 12 Weeks—Free Microsoft Course

Looking to build a strong foundation in machine learning? Microsoft’s “Machine Learning for Beginners” curriculum is a free, hands-on 12-week course designed to take you from basic concepts to practical applications. Whether you’re new to ML or seeking to refresh your skills, this resource offers a structured, project-driven path to mastery.

• Covers 26 lessons and 52 quizzes, spanning regression, classification, clustering, NLP, time series, and reinforcement learning—all using Python and Scikit-learn.

• Each lesson includes pre- and post-lecture quizzes, written instructions, step-by-step projects, assignments, and knowledge checks for active learning.

• Real-world datasets and global themes make the material engaging and relevant, with projects like price prediction, sentiment analysis, and recommender systems.

• Multi-language support (50+ translations) and optional R lessons ensure accessibility for learners worldwide.

• Community features: join the Microsoft Foundry Discord for collaborative learning, troubleshooting, and ongoing support.

Ideal for students, professionals, and educators, this curriculum is perfect for anyone wanting to learn classic machine learning through practical, guided projects. Dive in to accelerate your ML journey!

Tool of the week

whichllm — Auto-rank the best local LLMs for your hardware

Deploying large language models (LLMs) locally can be a guessing game: which model fits your hardware, and which one actually performs best? whichllm is an open-source CLI tool that benchmarks, ranks, and recommends the top LLMs from HuggingFace that will run efficiently on your specific system—no manual research or trial-and-error required.

• Auto-detects your hardware (GPU/CPU/RAM) and instantly ranks compatible models by real-world benchmarks, not just size.

• Evidence-based recommendations using live data from sources like LiveBench, Open LLM Leaderboard, and Chatbot Arena ELO, with recency and confidence grading to avoid stale or inflated scores.

• Flexible simulation and planning: Test with hypothetical GPUs (e.g., RTX 4090), multi-GPU setups, or plan hardware upgrades before you buy.

• One-command chat and code snippets: Instantly start a chat with the best model for your rig, or generate ready-to-run Python code for any supported LLM.

• Scriptable, filter-rich CLI: Output in JSON or Markdown for pipelines, filter by speed, fit type, task profile (coding, vision, math), and more.

whichllm is gaining traction among AI developers and hobbyists for its practical, evidence-driven approach. Check out the GitHub repo for install instructions and join the growing community (over 1,000 stars and counting).

Top News of the week

US Export Controls Force Anthropic to Disable Fable 5 and Mythos 5 Models Globally

In a dramatic escalation of AI regulation, the US government has ordered Anthropic to immediately suspend access to its most advanced AI models, Fable 5 and Mythos 5, for all users worldwide. The directive, issued under national security authorities, applies to any foreign national—regardless of location—and even includes foreign national Anthropic employees. This move marks an unprecedented intervention in commercial AI deployment, raising urgent questions about the reliability and continuity of AI services amid shifting geopolitical landscapes.

Anthropic stated that the government’s action was prompted by concerns over a potential “jailbreak” technique that could bypass Fable 5’s safeguards. However, Anthropic maintains that the vulnerabilities identified are minor, not unique to their models, and are present in other leading AI systems. The company emphasized its “defense in depth” strategy, including extensive red-teaming and robust safeguards, and expressed disagreement with the government’s decision, arguing that such a standard could halt frontier model deployments industry-wide. Access to other Claude models remains unaffected, but developers relying on Fable 5 and Mythos 5 must now transition to alternative solutions.

This incident underscores the growing influence of government regulation on AI innovation and highlights the importance for organizations to diversify their AI providers to mitigate risks from sudden regulatory actions.

🔗 Read the official announcement

GLM-5.2 Sets New Open-Source Coding Benchmark

GLM-5.2 has claimed the top spot on the DeepSWE leaderboard, outperforming previous leaders like Kimi K2 by a notable margin. With a 44% pass@1 score and a 1 million token context window, GLM-5.2 demonstrates robust bug-fixing capabilities on real-world open-source repositories—without internet access or hints. Released under the MIT license, it enables commercial use and self-hosting, empowering developers to build coding agents for practical, complex scenarios.

🔗 DeepSWE Leaderboard and Blog

Hermes Agent v0.17.0 Expands Automation and Messaging Features

Nous Research’s Hermes Agent v0.17.0 introduces major enhancements, including iMessage support via Photon Spectrum, background subagents for asynchronous tasks, and Automation Blueprints for natural language scheduling. The update also brings WhatsApp Business Cloud API integration, Raft agent network support, and improved desktop app functionality. With contributions from 245 developers and nearly 1,500 commits, Hermes continues to push the boundaries of agentic automation and communication.

🔗 Hermes Agent Documentation

Google Gemma 4 12B Coding Model Runs Locally on 12GB VRAM

The Gemma 4 12B coding model, now available through Hugging Face, can run locally on machines with as little as 12GB of VRAM. This makes advanced code generation and reasoning accessible to a wider range of developers, eliminating the need for cloud resources. The model features a 256K context window and is optimized for Python coding tasks, offering strong performance and flexibility for offline workflows.

🔗 Gemma 4 12B Coding Model on Hugging Face

VibeThinker-3B: Small Model Matches Frontier Reasoning Performance

VibeThinker-3B, a compact 3B parameter model, achieves frontier-level performance on verifiable reasoning tasks, rivaling much larger LLMs like GLM-5 and Gemini 3 Pro. It scores 94.3 on AIME26 and 80.2 Pass@1 on LiveCodeBench v6, and generalizes well to recent LeetCode contests. The report highlights that verifiable reasoning can be compressed into small models, making them efficient alternatives for demanding coding and math tasks.

🔗 VibeThinker-3B Technical Report

Tip of the week

Supercharge Your AI Projects with Specialized Agent Prompts

Struggling to get consistent, high-quality results from generic AI prompts? The open-source "Agency Agents" library offers 232+ production-ready, personality-driven AI agent prompts for every technical and business role.

• What it is: A curated collection of expert prompts—each agent has a unique voice, domain expertise, and proven workflow. Examples include Frontend Developer, Backend Architect, Growth Hacker, Incident Response Commander, and many more.

• How to apply: Install agents for your preferred tool (Claude Code, GitHub Copilot, Cursor, Gemini CLI, OpenCode, etc.) using simple scripts:

• Why it's useful: Agents deliver measurable outcomes, real code, and actionable advice—no more vague "act as a developer" prompts. Pick only the teams/divisions you need for your workflow.

• Key benefit: Streamline prompt engineering, boost productivity, and ensure reliable, specialized outputs for every project phase.

Try this when you need expert-level AI support for coding, design, marketing, project management, or testing.

Explore the Agency Agents library on GitHub

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help