Subscribe, if this was forwarded to you

Hey there,

This week’s edition is packed with breakthroughs and practical insights from the frontlines of AI and technology. We spotlight NASA’s historic use of Anthropic’s Claude AI to autonomously drive the Perseverance rover on Mars, marking a new era for robotic exploration. On the developer front, our featured tutorial walks you through running GLM-4.7-Flash locally for agentic coding workflows, while GLM-OCR emerges as a benchmark-topping open-source solution for multimodal document understanding. Plus, we dive into the latest survey results on AI-assisted coding adoption and highlight innovations like Qwen3-Coder-Next and Cursor’s secure codebase indexing.

Let’s dive in!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Run GLM-4.7-Flash locally for coding

🛠️ Tool Spotlight: GLM-OCR: open-source multimodal document understanding

📰 Top News: Claude AI powers Mars rover autonomy

💡 Tip: Survey reveals cautious AI code adoption

Let's get started!

Tutorial of the week

Run GLM-4.7-Flash Locally for Coding AI

Looking to supercharge your local coding assistant? This step-by-step guide from Pondhouse Data shows you how to run Zhipu AI’s powerful GLM-4.7-Flash model on your own hardware using llama.cpp and OpenCode. With open weights and an MIT license, you get full control—no API costs or data leaving your machine.

• Covers both NVIDIA (CUDA) and AMD (ROCm) GPUs, with clear build instructions for each platform.

• GLM-4.7-Flash is a 30B parameter Mixture-of-Experts model, but only ~3B parameters are active per pass—making it practical for 16GB+ GPUs.

• Benchmark results show outstanding coding performance, nearly tripling scores of comparable open models on SWE-bench and other agentic tasks.

• Walks you through downloading the model, setting up the server, and wiring it to OpenCode for agentic coding workflows.

• Includes troubleshooting tips, configuration examples, and links to further reading on self-hosted LLMs and cost-saving strategies.

This tutorial is ideal for developers, data scientists, and AI engineers who want a high-performance, local coding assistant without cloud dependencies. If you have a capable GPU and want to explore advanced agentic AI, this guide is a must-read.

Tool of the week

GLM-OCR — State-of-the-art open-source multimodal document understanding

GLM-OCR, developed by Z.ai, is a compact yet powerful multimodal OCR model designed for advanced document understanding. With just 0.9B parameters, it excels at parsing complex layouts, tables, and formulas, making it ideal for real-world business scenarios where document structure varies widely.

• Benchmark-leading performance: GLM-OCR ranks #1 on the OmniDocBench V1.5 with a score of 94.62, outperforming larger models in formula, table, and information extraction tasks.

• Efficient and scalable: Its lightweight architecture enables fast inference (up to 1.86 PDF pages/sec), supporting high-concurrency and edge deployments without sacrificing accuracy.

• Flexible integration: Supports deployment via vLLM, SGLang, Ollama, and Hugging Face Transformers, with a comprehensive SDK for easy integration into production pipelines.

• Robust real-world handling: Optimized for documents featuring complex tables, code snippets, seals, and challenging layouts, maintaining high recognition accuracy.

• Open-source and community-driven: Fully open-sourced under MIT, with over 96,000 downloads last month and 645+ stars on Hugging Face, plus active demos and integrations in the ecosystem.

GLM-OCR is rapidly gaining traction among developers and enterprises seeking reliable, efficient OCR for diverse document types. Explore the model and its SDK on Hugging Face.

Top News of the week

Anthropic’s Claude AI Powers First Autonomous Mars Rover Drive

NASA has achieved a major milestone in space exploration by using Anthropic’s Claude AI to autonomously plan and generate navigation commands for the Perseverance rover on Mars. For the first time, an AI system, rather than human engineers, plotted a 400-meter route across the Martian surface, marking a breakthrough in robotic autonomy and mission efficiency. This advancement could dramatically accelerate scientific discovery and reduce operational overhead for future planetary missions.

Claude’s involvement went far beyond simple route suggestions. NASA’s Jet Propulsion Laboratory (JPL) provided Claude with years of rover data and orbital imagery, enabling the AI to write executable navigation plans in Rover Markup Language, iteratively refine waypoints, and self-critique its work. The resulting plans were validated against simulations modeling over 500,000 variables to ensure safety and accuracy. Only minor human edits were required before the commands were transmitted to Mars, where Perseverance successfully executed the AI-generated path.

This successful test run demonstrates the potential for AI-driven autonomy in deep space missions, promising faster planning, more consistent operations, and greater scientific yield. As NASA looks ahead to ambitious projects like Artemis and beyond, AI assistants like Claude could become indispensable for exploring the Moon, Mars, and even more distant worlds.

🔗 Read the official announcement

Alibaba Unveils Qwen3-Coder-Next: Efficient Open-Weight Coding Agent

Alibaba’s Qwen team has released Qwen3-Coder-Next, a new open-weight coding agent model designed for efficient local deployment. Leveraging an 80B parameter Mixture-of-Experts (MoE) architecture, the model activates only 3B parameters at inference, dramatically reducing compute requirements without sacrificing performance. Trained on 800,000 executable coding tasks, Qwen3-Coder-Next excels at multi-step debugging, refactoring, and tool-driven coding workflows, setting new benchmarks on SWE-Bench.

🔗 Read the Qwen3-Coder-Next announcement

MIT Researchers Advance Continual Learning with Self-Distillation

A team from MIT has introduced On-Policy Self-Distillation Fine-Tuning (SDFT), a novel method for continual learning in AI systems. SDFT enables models to acquire new skills from demonstrations without explicit rewards and, crucially, without catastrophic forgetting of previous knowledge. Experiments show SDFT outperforms traditional supervised fine-tuning, especially as model size increases, paving the way for more robust, adaptable AI capable of sequential skill accumulation.

🔗 Explore the SDFT research

Anthropic Study: AI Coding Assistance May Impede Short-Term Learning

Anthropic’s latest randomized controlled trial reveals that while AI assistance can speed up coding tasks, it significantly reduces short-term conceptual understanding. Developers using AI to learn the Trio Python library scored 17 percentage points lower on immediate comprehension tests than those coding manually, particularly in debugging. The study highlights the need for intentional AI tool design to balance productivity with skill development in engineering teams.

🔗 Read the Anthropic study summary

Cursor Accelerates Codebase Onboarding with Secure Semantic Index Reuse

Cursor has dramatically improved agent onboarding speed for large codebases by enabling secure reuse of semantic indexes within teams. Using cryptographic Merkle trees and similarity hashes, Cursor allows new users to leverage existing indexes, reducing time-to-first-query from hours to seconds—even for repositories with tens of thousands of files. This innovation streamlines collaboration and enhances productivity for organizations working with AI-powered code agents.

🔗 Learn more about Cursor’s indexing breakthrough

Tip of the week

The AI Coding Reality Check: Widely Used, Carefully Checked

Curious about how your peers are adopting AI-assisted coding tools? A recent survey from SonarSource sheds light on the current landscape of AI in software development, revealing both enthusiasm and caution among developers.

• Daily usage: 72% of developers are using AI coding tools on a daily basis.

• Trust levels: Despite widespread use, only 4% of respondents fully trust the functional correctness of AI-generated code. The majority (96%) remain cautious, highlighting concerns about reliability and security.

• Common practices: Most developers review and test AI-generated code before merging or deploying, often using automated tools for static analysis.

• Challenges: The survey found that subtle bugs and potential vulnerabilities are top concerns, with many developers reporting that AI suggestions can look correct but may not be functionally reliable.

These findings reflect a community that is eager to leverage AI for productivity, but also aware of its limitations. As AI tools continue to evolve, developers are balancing innovation with careful review and quality assurance.

Read more: SonarSource AI Code Quality Survey

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