Runtime API: patch only. Upstream fixed an ABI mismatch in libGemmaModelConstraintProvider that blocked EmbeddingGemma 2 when building LiteRT-LM from source. No Kotlin API change. Kai consumes the Google Maven artifacts and does not build LiteRT-LM from source.
iOS: still on LiteRT-LM Swift package 0.16.1. Upstream tagged v0.18.1; moving iOS still needs the SPM pin and the Swift bridge ported to ModelInfo.
Runtime API: 0.18 removed Capabilities; its replacement is ModelInfo.from(path), which returns ModelInfo.Llm / Embedding / Unknown. Function-calling, thinking and sampler defaults now live on ModelInfo.Llm; input modalities on the base class. A non-LLM bundle reports unknown capabilities, so callers keep their own defaults as before.
iOS: still on LiteRT-LM Swift package 0.16.1. Upstream has since tagged v0.17.0, v0.17.1 and v0.18.0 on GitHub (each with Package.swift), so iOS can now follow — not done here; it needs the SPM pin moved and the Swift bridge ported to ModelInfo.
2026-09-07
Pin check: all four existing pins attested against the HuggingFace tree API at their pinned commits.
Bump — gemma-4-12b-it: upstream replaced the weight file on 2026-09-03 (161028364f39), size 6547589312 → 6883278368. Pinned to 7a0b1ce0ea821bcd01c5f72af84155e02191152f, sha256 58fd31b7…. The new build adds vision + audio modalities and Multi-Token Prediction, and requires litert-lm ≥ 0.17; Android/JVM moved to 0.17.0 on 2026-09-06, so the requirement is met. The iOS bridge (LiteRT-LM 0.16.1) cannot load it — accepted, since the file is 6.9 GB.
Added — lfm2.5-1.2b-instruct: LFM2.5-1.2B-Instruct_int4_gpu.litertlm at f45d8d8abe93bff4026efee20fa483150ce8e687, 736220768 bytes. Chosen over Qwen3.5 because its bundle carries a real tool-calling chat template (the Qwen3.5 conversions strip theirs). _int4_gpu rather than _int4 because only that build lowers fully for the GPU delegate while still running on CPU.
Runtime API: wired litert-lm 0.17's expanded Capabilities — declared function-calling support now gates whether an on-device model is handed tools at all, and each model's declared sampler defaults replace the hardcoded top-k 40 / top-p 0.95 / temperature 0.8.
iOS: LiteRT-LM Swift package 0.15.0 → 0.16.1 (the newest tag; 0.17.0 was never tagged). The Swift sources for Engine, Config, Conversation and Capabilities are byte-identical between the two tags, so only the prebuilt xcframework changes.
2026-08-12
Update: Live pin check via process:update-litert-models against HuggingFace models + tree APIs.
Attested — all four pins (gemma-4-e2b-it, gemma-4-e4b-it, gemma-4-12b-it, qwen3-0.6b) match .lfs.oid and .lfs.size at the pinned commit.
main moved — each repo sha is newer than the pin; the .litertlm LFS oid and size on main are identical to the pin. No bump (bytes unchanged; policy forbids silent pin moves).
Kotlin — unchanged.
Initialization: Created OKF litert bundle (pin policy + verify/bump playbook). Runtime source of truth remains LocalModelCatalog.kt.