v0.8.0 v0.28.0

high risk

msitarzewski/agency-agentsAug 25, 2026 → Aug 26, 20261 day between releases

What's New

Drastic shift from mouse configuration to LLM performance and model support

Breaking Changes

  • bitsandbytes support migrated to an out-of-tree plugin
  • Transformers bumped to version 5.15.0
  • Deprecated calculate_kv_scales runtime function removed
  • Deprecated override_attention_dtype removed

New Features

  • New model support: Muse Glimmer, Ling 3.0 Flash, Dots3 NOTE, and Interns2mobius
  • Vision encoder optimizations including ViT full CUDA graphs
  • New defaults: max_num_batched_tokens raised to 16384 and prefix caching enabled by default
  • RL workflow improvements including stateful trainer send over NCCL
  • Hardware support for AMD Quark NVFP4 and ROCm gfx11/gfx950

Improvements

  • Significant Kimi-K3 performance push with Decode Context Parallel (DCP) support
  • DeepSeek V4 sparse MLA support and AMD ROCm enablement
  • Advances in speculative decoding with DFlash2 and DSpark confidence-scheduled verification
  • Rust frontend and gRPC enhancements including multimodal image inference
  • Tiered KV cache offloading with disk support and performance metrics

Bug Fixes

  • Narrowed eager CUDA graph regions to optimize memory usage

Upgrade Notes

Migrate bitsandbytes to the new out-of-tree plugin and update Transformers to 5.15.0. Check for usage of the removed calculate_kv_scales and override_attention_dtype functions. New defaults include a raised max_num_batched_tokens and enabled prefix caching.

The Two Releases

v0.8.0

v0.8.0

Aug 25, 2026· 2 months ago

The major v0.8.0 release introduces significant GUI and functionality improvements. It unifies device settings across connection types, redesigns the device workspace for better scalability, adds per-app profile editing in the GUI, and enhances scrolling capabilities with smooth scrolling and tap/hold differentiation.

Full release notes
v0.28.0

v0.28.0

Aug 26, 2026· 2 months ago

A major performance update focusing on Kimi-K3 optimizations, DeepSeek V4 support, and speculative decoding advances, alongside Rust frontend and KV cache improvements.

Full release notes