AI / LLM

Patch release addressing a floating point exception crash in the gemma4:12b model and improving Windows installation for Hermes integrations.

Adds Cline CLI auto-installation, Qwen code integration, Radeon 8060S iGPU support, and Laguna architecture support. Includes security hardening for markdown URL handling, improved logging for troubleshooting, and various fixes for model loading and token counting.

Added support for the Gemma4-12B model, expanding the available model options for users.

Updates llama.cpp dependency and fixes Windows cleanup process by properly terminating llama-server. Known issue with gemma4:12b model causing crashes due to floating point exceptions.

Adds Codex App integration for desktop development workflows with built-in browser, code review features, and git support. Improves MLX sampler quality on Apple Silicon. Introduces new model recommendations for coding tasks.

Added vision model support with image inputs to ollama launch opencode command and fixed formatting issues with Claude tool results when using local image paths.

Patch release improving MLX model handling with refined push behavior, enhanced image generation runner thread affinity, fixed inference timeouts, and resolved macOS target leakage issues. Also includes test hardening and update flow improvements.

Adds Gemma 4 MTP speculative decoding support for Mac MLX runner providing 2x speed improvements for coding tasks, updates MLX libraries with threading fixes, and bumps Go version to 1.26.

Adds Claude Desktop integration with launch command, introduces server-driven model recommendations in the app, fixes Windows gateway timeout issue, and improves Metal initialization error handling.

Added support for two new models: NVIDIA's Nemotron 3 Omni and Poolside's Laguna XS.2 coding model.

Minor update improving Gemma 4 model renderer for thinking and tool calling, enabling dynamic model recommendations without app updates, aligning desktop launch page with CLI integrations, and fixing Poolside integration title display.

Release candidate adds support for 'max' as a think parameter value in the API and maps OpenAI response reasoning effort to think functionality.

Minor update improving OpenClaw onboarding reliability, standardizing model ordering in launch command, and bundling web search plugin with OpenClaw integration.

Adds Kimi CLI integration for agentic execution tasks, improves MLX runner performance with logprobs support and faster sampling, enhances thread safety and tokenization, fixes GLM4 MoE Lite router performance, and resolves UI issues in macOS app model picker and Gemma 4 structured outputs.

Introduces Hermes Agent for adaptive workflow assistance, adds Gemma 4 MLX support for Apple Silicon with mixed-precision quantization, enhances ollama launch with GitHub Copilot CLI integration and improved configuration handling, fixes OpenClaw non-interactive setup and various build issues includ

Patch release that improves quality for specific Gemma model variants when thinking mode is disabled and updates ROCm to version 7.2.1 on Linux systems.

Enhances Gemma 4 tool calling with Google's latest fixes, improves parallel tool calling for streaming responses, adds Hermes agent integration documentation, and resolves image attachment errors in the Ollama app.

Adds OpenClaw messaging channel integration via new launch command, enables flash attention for Gemma 4 models on compatible GPUs, improves OpenCode detection for curl-based installations, and fixes save command issues with safetensors-imported models.

Minor update enhancing Gemma 4 tool calling functionality, adding new models to the application interface, and resolving issues with OpenClaw TUI launcher.

Performance improvements for M5 chips using NAX optimization and enabled flash attention for Gemma4 model to enhance processing efficiency.

Changed default app home view from launch screen to new chat interface for improved user experience.

Patch release with benchmarking improvements, Gemma4 model parsing fixes for quoted strings and tool calls, GGML optimizations for CUDA batched operations, flash attention enablement for Gemma4, and ROCm build fixes.