Skip to content

Policy

A pin is attested when the HuggingFace tree API for the pinned commit reports the same SHA-256 and byte size as Kotlin. main moving is not a reason to bump if those bytes are unchanged. See pin-policy.md.

Replace this snapshot only via the refresh playbook. Never write /resolve/main/.

Current pins

Snapshot mirrored into MODEL_CATALOG (runtime). 5 models. GPU baseline and context defaults stay in Kotlin only.

Id Repo Pinned commit File SHA-256 sizeBytes Pin vs tree
gemma-4-e2b-it litert-community/gemma-4-E2B-it-litert-lm 9262660a1676eed6d0c477ab1a86344430854664 gemma-4-E2B-it.litertlm 181938105e0eefd105961417e8da75903eacda102c4fce9ce90f50b97139a63c 2588147712 attested
gemma-4-e4b-it litert-community/gemma-4-E4B-it-litert-lm f7ad3343bd6ebc9607f4dc3bc4f2398bd5749bc5 gemma-4-E4B-it.litertlm 0b2a8980ce155fd97673d8e820b4d29d9c7d99b8fa6806f425d969b145bd52e0 3659530240 attested
gemma-4-12b-it litert-community/gemma-4-12B-it-litert-lm 7a0b1ce0ea821bcd01c5f72af84155e02191152f gemma-4-12B-it.litertlm 58fd31b778ca2c21c80d634fb34fc5a89d11d563a38dfd3cbf1b40dbf252a8b6 6883278368 attested (bumped 2026-09-07)
lfm2.5-1.2b-instruct litert-community/LFM2.5-1.2B-Instruct f45d8d8abe93bff4026efee20fa483150ce8e687 LFM2.5-1.2B-Instruct_int4_gpu.litertlm 36f7f0221bcc42c75291da1d7e3422901024a5b06b9bfa3c02d7feface04f70a 736220768 attested (added 2026-09-07)
qwen3-0.6b litert-community/Qwen3-0.6B dd97997951bb15a2a71f539ba17f604707c0b11a Qwen3-0.6B.litertlm 555579ff2f4fd13379abe69c1c3ab5200f7338bc92471557f1d6614a6e5ab0b4 614236160 attested

main at the time of the check

Id main commit main file vs pin
gemma-4-e2b-it b3ca0d2f0767 same oid + size
gemma-4-e4b-it 2eee7ac325f2 same oid + size
gemma-4-12b-it 7a0b1ce0ea82 pinned to main (that is where the new bytes are)
lfm2.5-1.2b-instruct f45d8d8abe93 pinned to main
qwen3-0.6b 8414150f2e9d same oid + size

Snapshot

Field Value
Fetched 2026-09-07
Source HuggingFace models API + tree API at the pinned commit and at main
Attested 5 / 5 pins match tree oid + size
Bumped gemma-4-12b-it (new weights upstream, not a metadata move)
Added lfm2.5-1.2b-instruct

Notes

  • The gemma-4-12b-it bump is a real reweight, not a main drift: upstream commit 161028364f39 ("Updated the gemma-4-12B-it.litertlm file", 2026-09-03) replaced the file, and the byte size moved 6547589312 → 6883278368. The new build adds vision + audio modalities and Multi-Token Prediction, and its card requires litert-lm ≥ 0.17 — which the Android/JVM runtime now is. Users who already hold the old file re-download 6.9 GB.
  • lfm2.5-1.2b-instruct pins the _int4_gpu build deliberately. The plain _int4 file delegates 536 of 579 ops and then fails engine creation (GATHER_ND plus INT64 tensors do not lower), while _int4_gpu runs on both GPU and CPU — required, because initialize() tries GPU first and falls back.
  • Qwen3.5 (0.8B / 2B / 4B) exists in litert-community and was not added: those bundles ship a simplified ChatML template with the tool-calling and vision sections deliberately removed, which would repeat the Qwen3 0.6B problem.
  • Runtime download URLs stay on the pinned commits. A newer main sha with the same LFS oid is repo metadata, not a new weight file.
  • User imports are not in this catalog.