What's Changed
- Add LFM3 media capabilities and consume the family runtime (#3230) by @jjang-ai
- Let bundle agents stay out of the Orchestrator's delegation list (#3227) by @tpae
- deps: vmlx-swift 085c1c57 — DeepSeek-V4.1 F32 control parameters for every bundle (#3214) by @jjang-ai
- deps: pin vmlx-swift 1410c6f4 (DeepSeek-V4.1 JANGHT keeps F32 control parameters) (#3213) by @jjang-ai
🐛 Bug Fixes
- Save generation defaults per custom agent (#3233) by @jjang-ai
- Remove the Next turn control from the composer (#3232) by @jjang-ai
- Turn on reasoning for Anthropic, Claude Code, and OpenAI models from their live contracts (#3229) by @tpae
- Coalesce chat preview composition and memoize agent routing JSON (#3198) by @sentry[bot]
- Avoid main thread hang parsing router activity timestamps in Credits (#3177) by @sentry[bot]
- Fix AppHang during system appearance changes in ThemeManager.isDarkMode (#3183) by @sentry[bot]
- Move phone chat handoff system readings off main thread (#3221) by @sentry[bot]
- Offload isGitRepository filesystem probe from main actor in KnowledgeManager (#3222) by @sentry[bot]
- Use Foundation character-set search for newline checks in agent descriptions (#3197) by @sentry[bot]
- Preserve template variables and per-turn generation controls (#3226) by @jjang-ai
- Preserve saved settings and external model availability (#3228) by @jjang-ai
- Reflow source-wrapped paragraphs in model cards (#3219) by @jjang-ai
- Keep message content order through image preprocessing (#3224) by @jjang-ai
- DeepSeek-V4.1 alignment: dSpark solo lease, full effort ladder, model_type detection, cached-token usage (#3216) by @jjang-ai
- Track preparation failures and connect resident cache fixes (#3218) by @jjang-ai
- Unify installed model inventory and correct capacity estimates (#3217) by @jjang-ai
- Own allocator cleanup per request and retain failed MTP diagnostics (#3215) by @jjang-ai
🧰 Maintenance
- Pin vmlx-swift c6fee283: DeepSeek-V4.1 decode attention, dSpark governor fix, SSD restore past 16K tokens (#3223) by @jjang-ai
Full Changelog: 0.25.27...0.25.28
Downloads
- Osaurus-0.25.28.dmg — standard build. Pick a model during onboarding.
- Osaurus-0.25.28-full.dmg — 3.6 GB, Raptor 0.6 bundled. Chat offline right after install; no model download.
sha256 cb23748f86bf77ff156e9680e6de70c7b58fed8beac5456a67fa7a3349b09fb2