ENPT

d1-omni-600M

Last updated:

Open—$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity5.9/10

    Experimental open weights under the same LFM Open License v1.0 as d1-3B, with a README for text, image and audio. GGUF exists. Liquid says inference numbers are not ready. No hosted API or SLA.

  • Capability5.5/10

    Same noul/choice/score interface, read from the option distribution. No ECE and no latency table. Decision Index 15.95 is Liquid's own run. Audio is limited to English assistant clips of at most 30 seconds.

  • Adoption4/100

    About 8.5k downloads and 77 likes in three days, ahead of d1-3B on downloads, plus a GGUF repo. The launch attention sat on the 3B. No production case.

Vendor claims

d1-omni-600M is the smaller open checkpoint in Liquid AI’s 7 Oct 2026 release. The card lists 587M parameters: a 381M trunk and decision head, a 94M vision encoder and a 112M audio encoder. The base is LFM2.5-Encoder-350M, a bidirectional encoder, not the decoder used by d1-3B.

A state is text or JSON, with images or one voice clip. Answers come from the distribution over the options, with zero output tokens. It accepts either text and image or text and audio, not all three at once. Context is 16,384 positions. With images, the state and question text is cut to 896 tokens. Audio clips are cut at 30 seconds. transformers>=5.15, trust_remote_code=True.

License. Same LFM Open License v1.0 as d1-3B. Commercial use is limited once annual revenue reaches $10M.

Audio scope. The card warns that audio was trained on requests between an English speaker and an assistant: utterance type, topic, and what the speaker wants.

What Liquid says, and what it does not

  • Decision Index v0.2.1: 15.95, against 48.57 for d1-3B. UnverifiedClaim.
  • Seven text benchmarks, mean 78.4, just above Decider 2B (77.1). Highest in that table on Civil Comments (95.8) and PAWS-X (79.5). UnverifiedClaim.
  • The blog does not publish a latency table. It calls this an early research release under active development. The 0–0 ms frontmatter means no figure.
  • No ECE.

The training write-up: fine-tune the encoder on decision tasks, then add audio (FastConformer plus an adapter, encoder trained with the text trunk frozen) and vision (SigLIP2 tower from LFM2.5-VL-450M, adapter plus LoRA that is active only on image inputs, vision encoder frozen). A later full fine-tune merges the LoRA and averages weights with the previous checkpoint.

Why it is in the catalog

It has its own trained decision weights and the same typed interface as d1-3B. The index score is far below the 3B, so it is a footprint option, not a second flagship. Compare d1-3B and the hosted d1.

Core ML port (8 Oct 2026)

FluidInference/d1-omni-600m-coreml converts the text path (LFM2.5 encoder trunk plus the decision head) to Core ML for Apple devices. The porter reports 2 answer flips in 492 checks against the original (max |Δp| 0.017) and about 4.6 ms per question at 64 tokens on the Neural Engine (UnverifiedClaim). It keeps the LFM Open License v1.0: no commercial use for entities with $10M or more in yearly revenue. Image and audio paths are not ported. Scores unchanged.