On June 16, 2026, developer Georgi Gerganov, creator of the ggml project, confirmed he uses the 27-billion-parameter open-weight Qwen3.6-27B model for daily production code maintenance on the ggml.org codebase. The disclosure, reported by Simon Willison, marks a verified real-world example of professional developers adopting local-llms for routine production coding work.
Georgi Gerganov Uses Local-LLMs Daily for ggml Production Maintenance
Gerganov’s confirmed daily use of Qwen3.6-27B for production code maintenance eliminates prior speculation about the viability of local coding models for professional work. The 27B parameter count of the Qwen3.6-27B model balances performance and hardware requirements, making it feasible to run on consumer-grade GPUs for full workday use cases. His workflow on the ggml.org codebase, a widely used open-source library for machine learning model inference, includes tasks such as bug fixes, feature implementation, and code refactoring, all performed without relying on cloud-based coding assistants. This use case aligns with growing developer interest in offline, privacy-preserving coding tools that do not require sending proprietary code to third-party servers.
GitHub Copilot Isolates Parallel AI Agent Sessions With Git Worktrees
GitHub’s Copilot blog published a breakdown of Git worktrees, a native Git feature that lets developers maintain multiple checked-out working directories for a single repository simultaneously. Each worktree is linked to a separate Git branch, eliminating the need to stash in-progress work or switch branches repeatedly when moving between tasks. For parallel AI coding agent workflows, this isolation prevents context loss and accidental overwrites between concurrent agent sessions. Traditional Git workflows that rely on stashing or frequent branch switching create clear risks: stashing in-progress changes can lead to lost work or merge conflicts when changes are reapplied, while frequent branch switching causes editor context loss as developers must close and reopen files or lose unsaved changes. Maintaining separate full repo clones for each task wastes disk space and complicates cross-branch changes. The GitHub Copilot desktop app creates a new Git worktree for each parallel AI agent session by default, keeping each agent’s work isolated from the developer’s active editor context and other agent sessions, per GitHub’s official Git worktree documentation.
Git worktree workflows have documented trade-offs. Each active worktree requires its own installed project dependencies, so disk usage increases with the number of concurrent worktrees. Stale worktrees should be removed with the git worktree remove command when no longer needed, and a single Git branch cannot be checked out in two worktrees at the same time.
GitHub’s Copilot CLI provides terminal-native slash commands to manage AI coding sessions across worktrees, reducing context switching overhead. Documented commands include /model to switch between AI models mid-session, /context and /compact to inspect and summarize token usage without losing the conversation thread, and /resume to restart a prior agent session by name.
For example, a developer running a long-form code generation agent can use /compact to summarize conversation history when the token budget nears its limit, preserving core context without losing work progress.
Additional commands support code review and permission management: /diff lets developers review every file change made by an agent before committing changes to the repository, /cwd lets users pivot the agent to a different repository without restarting the session, and /reset-allowed-tools lets developers drop agent permissions when moving to sensitive codebases, reducing the risk of unintended changes to critical code, per GitHub’s Copilot CLI command reference.
Allen AI Releases olmo-eval for Rigorous Local Model Testing
The Allen Institute for AI released olmo-eval, an evaluation framework built for the model development loop rather than just final leaderboard benchmarking. The framework extends the Open Language Model Evaluation Standard (OLMES) with features designed for rigorous, regression-focused testing of language models. It is intended for teams that need to validate model and prompt changes against specific task requirements, rather than relying on generic public benchmark scores. Key capabilities of olmo-eval include per-prompt comparison across model checkpoints, which lets teams catch performance changes on individual tasks rather than only seeing aggregate score shifts. The framework also includes minimum detectable effect reporting, which ensures observed score differences are statistically meaningful rather than random noise from benchmark variance. It uses modular runtimes with lightweight direct inference for question-answering benchmarks and containerized environments for code-execution tasks, reducing overhead for common evaluation workflows. Swappable components including judge models, tool sets, and prompt templates can be reused across different evaluation harnesses, cutting down on duplicate work for teams running regular model tests, per Allen AI’s official olmo-eval repository.
