ENPT

Liquid AI ships open d1-3B and d1-omni-600M

Last updated:

Source

On 7 Oct 2026 Liquid AI published two open-weight decision models. The hosted d1 API from 29 Sep is unchanged: its checkpoint is still unpublished, and these cards do not claim to be that model.

What shipped

  • d1-3B (3.12B), trained from LFM2.5-VL-3B. Text and images. noul, choice and score in one forward pass, output_tokens 0. Hugging Face, plus GGUF and an int8 build. Card created 5 Oct; the blog is dated 7 Oct.
  • d1-omni-600M (587M), trained from the LFM2.5 encoder. Text plus image, or text plus up to 30 seconds of audio, not all three together. Liquid calls it an early research release and does not publish a latency table.

Both use the LFM Open License v1.0. Commercial use is limited once annual revenue reaches $10M. The documented Python path is model.system_one(), with trust_remote_code=True, not a hosted /v1/systemone route.

Numbers to read carefully

Liquid reports Decision Index v0.2.1 public-split scores of 48.57 (d1-3B) and 15.95 (d1-omni-600M), and says 48.57 leads every model under 10B and matches Decider 35B-A3B (47.11). A seven-benchmark text mean is 82.9 for the 3B and 78.4 for the 600M. Single-question latency for the 3B starts at 8 ms on an RTX 4090 and 50 ms on a Jetson Orin Nano. All of that is UnverifiedClaim. The official Decision Index snapshot this site last checked (28 Sep) predates the cards.

@atomic_chat_hq (7 Oct, 17:28 BRT) ran d1-3B on one RTX 3090 in a Snake loop: 8 ms, 82 moves per second, 86 apples, against OpenAI’s Decisions API at 345 ms and 3 apples. The homepage at atomic.chat does not document the run. The comparison mixes a local forward pass with a network API. UnverifiedClaim.