What's Changed
- Support Qwen-Image-2.1 generation and editing (#2997) by @jjang-ai
- Restyle chat tab strip as a Safari-style track (#2995) by @tpae
- Pin Qwen4Exp JANGH runtime and media cache fixes (#2991) by @jjang-ai
- MCP robustness, elicitation, and regulated-industry connectors (#2990) by @tpae
- Use exact Naive rotary application on qualified M5 Max (#2972) by @jjang-ai
🐛 Bug Fixes
- Pin GLM sparse-indexer device range optimization (#3001) by @jjang-ai
- Pin canonical Qwen stable-prefix cache reconstruction fix (#3000) by @jjang-ai
- Stream Codex one-shot requests so context compaction works on ChatGPT models (#2999) by @tpae
- Render agent avatars at exact pixel size so they stay crisp (#2998) by @tpae
- Serialize image producers and honor completion boundaries (#2996) by @jjang-ai
- Pin native MTP history and sampled verification fixes (#2994) by @jjang-ai
- Pin runtime mixed-gather dtype correction (#2986) by @jjang-ai
- Avoid storage-unlock scheduler waits and stale starts (#2984) by @jjang-ai
Full Changelog: 0.25.17...0.25.18
Downloads
- Osaurus-0.25.18.dmg — standard build. Pick a model during onboarding.
- Osaurus-0.25.18-full.dmg — 3.6 GB, Raptor 0.6 bundled. Chat offline right after install; no model download.
sha256 9f521bccfce22df8ce9b8409ae5558d3137eeca609fadba7c2bf575bedb07fd3