desktop_v1.4.27 v0.30.0
high risk
minitap-ai/mobile-useSep 25, 2026 → Sep 25, 20260 days between releases
What's New
Major overhaul adds full-duplex framework and expanded NPU support
Breaking Changes
- diffusion_batch_size has been removed; use max_num_seqs instead
New Features
- Unified full-duplex serving framework (DuplexOmni, DuplexOmniEngine)
- Native cross-stage KV transfer with NIXL transport
- Interactive world-model and video execution (LingBot World)
- RL rollout serving API
- Multi-process API serving via --api-server-count
Improvements
- Engine-owned sessions separate duplex orchestration from turn-based execution
- Native Mooncake AR-to-DiT handoff and KV prefetch for diffusion models
- Realtime minimax-h3 4-step inference on Nvidia Blackwell and Ascend NPU 950
- Expanded model support including MAGI-2 and Ming-Image
- Improved batching, memory efficiency, and deployment scalability
Upgrade Notes
Replace the deprecated `diffusion_batch_size` configuration with `max_num_seqs`. The unified full-duplex framework changes session management, requiring updates to orchestration logic. Verify compatibility with new model plugins like MiniCPM-o 4.5 and AURA.
The Two Releases
v0.30.0
v0.30.0
Sep 25, 2026· 26 days ago
A major release featuring a unified full-duplex serving framework, native cross-stage KV transfer, interactive world-model and video generation, and expanded support for Nvidia Blackwell and Ascend NPU 950.
Full release notes