Release v3.0.3

🚀 Features

• Support Qwen 3.5 models (PR: #391)
• Upgrade LlamaLib to v2.0.5 (llama.cpp b8209) (PR: #391)
• add button to redownload LlamaLib (PR: #393)

Release v3.0.2

🚀 Features

• Cache LlamaLib to prevent re-downloads (PR: #386)
• Implement strategies for context overflow (chat truncation, chat summarization) (PR: #384)
• Upgrade LlamaLib to v2.0.4 (PR: #384)
• Re-introduce UI dropdown for level of debug messages (PR: #384)

🐛 Fixes

• Fix context overflow with caching and overflow strategies (PR: #384)
• Ensure macOS build includes the required runtime library (PR: #382)
• Fix inference for AMD GPUs using Vulkan (PR: #384)

Release v3.0.1

🚀 Features

• add Unity.Nuget.Newtonsoft-Json in the assembly definition (PR: #379)
• Update LlamaLib to v2.0.2 (llama.cpp b7777) (PR: #380)

🐛 Fixes

• fix running in Editor with Android/iOS platform selected (PR: #378)

Release v3.0.0

Rewrite the LLM backend, LlamaLib, as a standalone C++/C# library and adaptation of LLMUnity.

🚀 Features

• implement LlamaLib as object-oriented C++/C# library
• update llama.cpp to b7664
• fix Vulkan GPU backend
• Android 16kb support
• fix iOS - xcode building
• fix RAG functionality for iOS
• polish samples
• optimise streaming functionality and implement callbacks on the C++ end
• remove chat templates from LLMUnity, use the llama.cpp templating
• implement property checks
• common handlng for both json and gbnf grammars
• simplify integration of tinyBLAS (light GPU backend for Nvidia GPUs)
• transition of client / server functionality in LlamaLib

Release v2.5.2

🚀 Features

• Support Android x86-64 architecture (Magic Leap 2) (PR: #344)
• Combine ARM and Intel architectures of macOS (PR: #345)

Release v2.5.1

🚀 Features

• Allow JSON schema grammars (PR: #333)
• Add support for Qwen3 models (PR: #335)
• Add support for BitNet models (PR: #334)
• Upgrade LlamaLib to v1.2.5 (llama.cpp b5261) (PR: #335)

🐛 Fixes

• Fix Unity Editor hanging after stopping a completion and restarting scene (PR: #335)

Release v2.5.0

🚀 Features

• VisionOS support (PR: #299)
• Add support for Gemma 3 and Phi 4 models (PR: #327)
• Fix Android support for older devices (use ARMv8-A instead of ARMv8.4-A) (PR: #325)
• Upgrade LlamaLib to v1.2.4 (llama.cpp b4969) (PR: #325)
• Default number of predicted tokens (num_predict) to infinity (-1) (PR: #328)

Release v2.4.2

🚀 Features

• Integrate DeepSeek models (PR: #312)
• Update LlamaLib to v1.2.3 (llama.cpp b4688) (PR: #312)
• Drop CUDA 11.7.1 support (PR: #312)
• Add warm-up function for provided prompt (PR: #301)
• Add documentation in Unity tooltips (PR: #302)

🐛 Fixes

• Fix code signing on iOS (PR: #298)
• Persist debug mode and use of extras to the build (PR: #304)
• Fix dependency resolution for full CUDA and vulkan architectures (PR: #313)

Release v2.4.1

🚀 Features

• Static library linking on mobile (fixes iOS signing) (PR: #289)

🐛 Fixes

• Fix support for extras (flash attention, iQ quants) (PR: #292)

Release v2.4.0

🚀 Features

• iOS deployment (PR: #267)
• Improve building process (PR: #282)
• Add structured output / function calling sample (PR: #281)
• Update LlamaLib to v1.2.0 (llama.cpp b4218) (PR: #283)

🐛 Fixes

• Clear temp build directory before building (PR: #278)

📦 General

• Remove support for extras (flash attention, iQ quants) (PR: #284)
• remove support for LLM base prompt (PR: #285)