Termux Launcher v0.2.24 — Termux edition (com.termux)
On-device text embeddings, end to end.
Added
- EmbeddingGemma 300M (LiteRT) and Qwen3 Embedding 0.6B (MNN) in the model catalog.
- MNN-format text embeddings — new native embedding runtime over MNN's
Transformer::Embedding, exposed through/v1/embeddingsand/api/embed(OpenAI- and Ollama-compatible), alongside the existing LiteRT embedding path.
Fixed
- LiteRT embedding models now reliably fetch their SentencePiece tokenizer sidecar on download/import (handles
?download=URLs andsentencepiece.model/tokenizer.model/spiece.modelnaming), so embeddings work out of the box instead of failing withembedding_tokenizer_missing. - MNN embedding packages download their full weight set (including
.bintables such asembeddings_int4.bin).
Verified on-device: LiteRT embeddings (768-dim), MNN embeddings (1024-dim), and MNN chat (no regression).
Install all three APKs from the same pack (Main + API + Styling). Debug-signed (testkey).