Atomic Chat Files

🚀 New Features

New model support on llama.cpp

• Qwen 3.8 27B — native multimodal dense model, 262K context extendable to 1M via YaRN, Apache 2.0
• Muse Glimmer 30B — Meta's 30B agentic model, runs on a single GPU, with vision built in via a dedicated perception encoder
• Nemotron 3.5 Lightning — 30B MoE, 3B active, 1M context, Mamba-2 hybrid, built for high-volume steps in always-on agents
• Nanbeige4.2 3B — 3B agentic model that beats Qwen3.5-9B and Gemma4-12B on tool use and code-agent tasks

Other

• Reasoning effort levels are now in the UI for models that support them
• The Models section got rebuilt — easier to browse and pick from, so recommended models are easier to find
• ROCm on Windows — compatible AMD GPUs can now run local backends there too

🔧 Improvements & Fixes

• llama.cpp updated to the latest upstream
• Setup now auto-starts the smallest model it finds on your disk — one less step before your first chat
• Reasoning gets its own block above the activity blocks, with a clearer trigger icon
• Saved backends always show up in the llama.cpp dropdown
• Clearer model-load errors — separated by cause and properly localized
• 10+ smaller UI and stability fixes across onboarding, settings, and model handling

🙏 Contributors

Thanks to @stevibe, @worthant, @Vect0rM for their contributions to this release!