| Name | Modified | Size | Downloads / Week |
|---|---|---|---|
| Parent folder | |||
| README.md | 2026-10-30 | 1.3 kB | |
| v.1.3.5 source code.tar.gz | 2026-10-30 | 27.7 MB | |
| v.1.3.5 source code.zip | 2026-10-30 | 28.1 MB | |
| Totals: 3 Items | 55.8 MB | 2 | |
LFM 3 is here
Android 173, iOS 26.
- LFM 3 - 1.7B: reasons before it answers, in 16 languages (#499)
- LFM 3 VL - 2B: ask about your photos (#499)
- The Think switch turns LFM 3 reasoning on or off; the think block streams from the first token and collapses when the think phase completes (#500)
- Steadier chats: a turn that produces nothing for five minutes stops and offers a retry, and earlier reasoning is no longer sent back with the next question (#500)
- Mixed-language conversations: a Latin-script question is answered in its own language even after a turn in another script (#501)
- Voice input: a silence marker from Whisper is no longer treated as something you said (#501)
- react-native-executorch 0.11.0; multimodal prompts are embedded in
max_seq_lenchunks, which LFM 3 VL needs to answer about a photo at all (#498) - The model catalog is now gated by app version: LFM 3 appears only in
model-catalog-v2.json, read from 1.3.5 on, while older installs keep the frozenmodel-catalog.json(#503, [#504])
Speeds per device and the full dev-benchmark table are in docs/BENCHMARKS.md (github.com).
Release branch: [#502].
Full Changelog: https://github.com/software-mansion-labs/private-mind/compare/1.3.4...1.3.5
