MaIN.NET.10.2.0

10.2.0 release

• Fixes local tool calling on the Self backend — messages are now added to the chat template on every iteration of the tool loop, preserving conversation context across tool calls.
• Upgrades LLamaSharp dependencies to 0.27.0 (LLamaSharp, LLamaSharp.Backend.Cuda12, LLamaSharp.kernel-memory).
• Expands the local model catalog with tool-calling optimized models: Qwen 2.5 7B Instruct, Llama 3.2 1B Instruct, Mistral 7B Instruct v0.3, Qwen 3.5 4B/9B (Unsloth), NVIDIA Nemotron 3 Nano 4B, and IBM Granite 4.1 3B.
• Adds model-specific tool-call format detection and prompt generation supporting Hermes/Qwen2.5, IBM Granite, Llama 3.1+, Mistral v0.3+, and Phi-3.5 native formats.
• Fixes hosted tool execution loop — server-side tools (web_search, get_current_datetime, etc.) now return final responses instead of hitting the max iteration cap.
• Makes InferPage's OpenAI-compatible API recognize both OpenAI hosted tool formats: {"type": "web_search"} and {"type": "web_search_preview"} (in addition to the existing MaIN function format).
• Adds new cloud models: Claude Fable 5, Claude Mythos 5, Claude Opus 4.8, Claude Sonnet 5, OpenAI GPT-5.6 Sol/Terra/Luna.
• Comprehensive documentation updates — new Supported Models reference and InferPage OpenAI API reference for model discovery and migration.
• Adds SearXNG integration for hosted web search — self-hosted meta-search engine that aggregates results from multiple search engines without API keys or rate limits. Configure via MaIN__SearxngBaseUrl environment variable or MaIN:SearxngBaseUrl in appsettings.json. Falls back to DuckDuckGo HTML scraping if not configured.