d1-Omni-2B
Released Oct 27, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- United States
- Type
- System One
- Modality
- Multimodal
- Context window
- 128K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Strengths & weaknesses
Strengths
- Takes the question itself as speech as well as text, so a voice request can be routed or checked without a transcription step in between
- Keeps most of its accuracy when questions are heard rather than read, according to Liquid's spoken-question runs
- Batches up to 16 written or spoken questions about one state into a single call, so a call-triage pipeline asks everything at once
- Liquid ships ONNX, GGUF and MLX builds, with a per-decision time on an Apple laptop chip in the tens of milliseconds
Weaknesses
- Cannot produce a transcript or speech — Liquid's card is explicit that it is not a chat model, so it only decides
- Spoken questions don't cover Hindi, Thai, Vietnamese or Indonesian, though those languages work in text
- Audio has to arrive at 16 kHz, and adding a ten-second clip roughly doubles per-decision time on a datacentre GPU
Not an LLM: the audio-capable open-weights "System One" decision model from Liquid AI's LFM3 family, 2.17B parameters built on LFM3-Omni-2B, taking text with images or audio as state plus written or spoken typed questions and returning calibrated probabilities in one forward pass, so none of this site's benchmarks apply and every cell is null by design. The model card gives a 128K-token context (recorded as 128,000; the exact integer was not seen), DecisionBench v1 76.4 and decisionbench.ai v4 84.7 (against 77.1 and 85.2 for d1-2B), 83.4 written vs 78.2 heard on the v1 routing-and-triage subset, and 5.8 ms per decision on an H100 (12.4 ms with a 10-second clip; 21.1 ms on an Apple M5 Pro); all are Liquid's own figures. Liquid makes no replacement claim for the earlier d1-omni-600M, which its docs still list, so predecessorId is null. Availability is self-host: open weights only, with no hosted endpoint. No price, knowledge cutoff or Artificial Analysis entry. Sources are search excerpts of liquid.ai/blog/lfm3, huggingface.co/LiquidAI/d1-Omni-2B (and its ONNX card) and docs.liquid.ai/lfm/models/decision-models, not pages read directly.
Benchmarks0 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
Not researched
Licence
LFM Open License v1.0 · Restricted · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor