Release 1.26.4

## What's Changed:
Version 1.26.4 fixes ggml backend loading issues that caused duplicate device name errors when both kronk and bucky SDKs are initialized in the same process, and updates the default llama.cpp library version from b9334 to b9360.

Release 1.26.3

## What's Changed:
Version 1.26.3 introduces FFmpeg-based audio transcoding support for the Bucky SDK, enabling transcription of additional audio formats beyond WAV, FLAC, and MP3. It also includes HTTP/2 improvements for peer communication, MCP service stateful mode, and various naming corrections.

Release 1.26.2

Release 1.26.1

## What's Changed
Version 1.26.1 brings support for Whisper models through the new Bucky audio transcription feature, along with CLI reorganization, BUI enhancements, pool engine refactoring, and updated documentation.

Release 1.26.0

## What's Changed
Version 1.26.0 brings significant improvements to MTP (Multi-Token Prediction) functionality, fixing critical bugs that prevented MTP from working correctly with IMC cache hits and hybrid models. It also fixes Anthropic API compatibility issues in the messages streaming path, ensuring correct input token counts and finish reasons are reported.

Release 1.25.9

## What's Changed
Version 1.25.9 introduces Multi-Token Prediction (MTP) speculative decoding, enabling faster inference with models that have built-in MTP heads — no separate draft GGUF required. This release also includes download workflow improvements, progress logger fixes, and comprehensive documentation updates.

Release 1.25.8

## What's Changed
Version 1.25.8 brings a yzma library upgrade that resolves heap corruption issues, adds local model file verification to prevent partial downloads from appearing fully installed, and fixes the BUI download resume flow so users can recover cancelled or truncated pulls.

Release 1.25.7

## What's Changed
Version 1.25.7 brings improved Mistral model support with reasoning effort handling, more reliable split-model downloads with size verification, and a cleaner HuggingFace pull experience in the BUI.

Release 1.25.6

## What's Changed
Version 1.25.6 brings a resource management fix and documentation updates.

Release 1.24.5

## What's Changed:
Version 1.25.5 brings bug fixes for VRAM calculation on models without GQA (such as BERT encoders) and improves RAM headroom defaults for small hosts.