AI Digest: Effective Agents, ZGCM-1, and Local GPUs

Today: an open 7B model for mathematics and agentic search, ways to make web agents cheaper, and fresh observations on running models locally.

🔥 Hot:

🔹 ZGCM-1 released as a fully open 7B model for mathematics and agentic search — The model combines internal reasoning with the use of external tools and focuses on training efficiency. 🔹 Researchers propose reducing web agents’ token usage by modeling application behavior — The approach is aimed at completing tasks in web interfaces without constantly analyzing every UI detail.

➡️ Useful materials:

🔹 OrchSLM explores how to orchestrate small language models instead of cloud LLMs — This approach could reduce latency, cost, and connectivity requirements in agentic systems. 🔹 A new study examines root-cause analysis for failures in long-running agent tasks — The authors propose looking for the cause of failure in extensive execution logs rather than limiting the analysis to the final result. 🔹 LabAgent helps AI agents preserve and develop the workflows of scientific laboratories — The idea addresses the problem of losing methods and context when laboratory staff change. 🔹 A study tests whether LLM agents can manage long-running physical tasks — The agent must continuously observe the environment, choose actions, and adapt to changes.

➡️ Discussions and case studies:

🔹 A user compares two AMD Radeon AI Pro R9700 cards with a pair of NVIDIA RTX 3090s for local models — The goal of the build is to run 30B models in FP8 or 70B models in Q4. 🔹 A user complains that a local coding harness gets stuck in a loop and forgets the task — Even when working on a simple web application, the model makes UI errors, so the author compares the experience with Claude Code.

📝 If you would like to add other news and materials to the list, write in the comments.