d1-2B
Released Oct 27, 2026 · knowledge cutoff unpublished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
- A third smaller than the d1-3B it replaces while answering more accurately on Liquid's own decision suites, so the same edge box handles more traffic
- Reads video as well as text and images in the state, so a recorded clip can be classified without pulling frames out first
- One forward pass with no generated tokens returns a calibrated probability per option, so a confidence threshold can decide what escalates to a larger model
- Ships in GGUF, MLX and ONNX builds from Liquid itself on day one, so it runs on laptops and phones without waiting for community ports
Weaknesses
- Cannot write text at all — it is a decision model, so anything that needs an explanation or a draft has to go to a generator
- Every accuracy figure so far is Liquid's own, on suites it chose; no independent replication had appeared at launch
- Free commercial use stops at US$10M annual revenue under the LFM Open License, and above that production use needs a paid licence from Liquid
Not an LLM: an open-weights "System One" decision model from Liquid AI's LFM3 family, 1.92B parameters built on LFM3-VL-2B (24-layer tri-hybrid backbone of gated short convolutions, sliding-window and global attention), taking text, image or video state plus 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. Liquid's LFM3 launch post says it "replaces our previous d1-3B, with a third fewer parameters", hence predecessorId. The model card gives a 128K-token context (recorded as 128,000; the exact integer was not seen), Decision Bench (decisionbench.ai) v4 85.2 against 83.6 for d1-3B, ImajevBench v2.0-lite 75.5 against 64.0, and 5.7 ms per decision on an H100 (3.8 ms on an MI325X); all are Liquid's own measurements. Availability is self-host: Liquid's docs list it as open weights run on your own hardware, and its hosted decision API serves d1-ultra and d1.1 instead. No price, knowledge cutoff or Artificial Analysis entry. Sources are search excerpts of liquid.ai/blog/lfm3, huggingface.co/LiquidAI/d1-2B 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
Replaces d1-3B