d1-3B
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity6.4/10
Weights on Hugging Face under the LFM Open License v1.0 (commercial use capped at $10M revenue, not an OSI license), with a README for noul/choice/score. Day-one GGUF and llama.cpp. No hosted API or SLA for this checkpoint.
- Capability7.3/10
Noul, choice and score in one pass, zero output tokens. The card says calibrated and publishes no ECE. The latency table and Decision Index 48.57 are Liquid's own runs, not the official board.
- Adoption5/100
About 5.3k downloads and 189 likes in three days, a GGUF repo, an HF arcade Space, and a third-party Snake demo (~1.9k likes). No production case yet.
Vendor claims
- Decision Index v0.2.1 public split 48.57, ahead of every model under 10B and level with Decider 35B-A3B at 47.11[Vendor claim — not independently verified]
- 8 ms per question on an RTX 4090, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin, 50 ms on Jetson Orin Nano, 30 ms on Apple M5 Pro[Vendor claim — not independently verified]
- Mean 82.9 on seven text benchmarks, ahead of Decider 4B at 81.1; vision score 74.1 vs LFM2.5-VL-3B 73.9 on 11 image benchmarks[Vendor claim — not independently verified]
d1-3B is Liquid AI’s open-weight decision model at 3.12B parameters, announced 7 Oct 2026. It is trained from LFM2.5-VL-3B (decoder-only VLM; the blog says the starting point was an average of LFM2.5-2.6B and the VL text backbone). It is not the hosted d1 API. That checkpoint is still unpublished, and these cards do not say they are the same weights.
You pass a state (text, JSON, images, or images alone) plus named questions. system_one() returns typed answers in one forward pass. usage.output_tokens is 0. Load with transformers>=5.14 and trust_remote_code=True. Context is 32,768 tokens. The vision encoder is SigLIP2 NaFlex, about 400M of the 3.12B.
License. Hugging Face lists
license: other, namelfm1.0. The file is the LFM Open License v1.0. Commercial use is limited once annual revenue reaches $10M. This is not Apache-2.0.
Specs
| Attribute | Value |
|---|---|
| Checkpoint | LiquidAI/d1-3B (card created 5 Oct 2026; blog 7 Oct) |
| Also | d1-3B-GGUF, d1-3B-w8a8 |
| Parameters | 3.12B (card) |
| Question types | noul, choice (named options), score (2–10 ordered levels, answer is the expected level) |
| Modalities | Text and images |
| Serving | In-process system_one / system_one_batch. Blog says llama.cpp from day one, including NVFP4 |
| Price | Free to download. No Liquid-hosted endpoint for this id |
Numbers to read carefully
All of these are Liquid’s. None of them is a row we checked on the official Decision Index Space (the public board we last used is the 28 Sep 2026 snapshot, which predates this checkpoint).
- Decision Index v0.2.1, public split: 48.57. The blog says that is ahead of every model under 10B and level with Decider 35B-A3B (47.11). UnverifiedClaim.
- Seven text benchmarks, mean 82.9 vs Decider 4B 81.1 and Decider 2B 77.1 (SQuAD 2.0, Civil Comments, MASSIVE, PubMedQA, BoolQ, XNLI, PAWS-X). UnverifiedClaim.
- Latency, one question, one request at a time: 8 ms RTX 4090, 9 ms MI325X, 30 ms Apple M5 Pro, 16 / 26 / 50 ms on Jetson AGX Thor, AGX Orin 64 GB and Orin Nano. Three questions on the 4090 are 21 ms. UnverifiedClaim. The frontmatter range 8–50 ms is that single-question table, not a measurement by this site.
- Vision: the blog does not report the private vision split of Decision Index v0.3. It says d1-3B scores 74.1 on 11 public image benchmarks, against 73.9 for the LFM2.5-VL-3B backbone. UnverifiedClaim.
A third party, @atomic_chat_hq (7 Oct 2026, 17:28 BRT), ran d1-3B locally on one RTX 3090 inside a Snake demo: 8 ms per call, 82 moves per second, 86 apples, against OpenAI’s Decisions API at 345 ms and 3 apples. The 8 ms matches Liquid’s 4090 figure and was measured on a different GPU. The apple count compares a local forward pass with a network API. UnverifiedClaim, and it is not a quality benchmark.
Fit / anti-fit
Fit when you need: a local multimodal decision model with the same question types as Jev, edge latency, or a checkpoint you can fine-tune.
Anti-fit when you need: a drop-in for the hosted d1 or for /v1/systemone over HTTP (the documented path is model.system_one in Python), an OSI license, or an accuracy number from a board this site has checked.
Why it is in the catalog
Liquid trained decision weights on top of its own VLM and published them. The interface returns a probability for each option (noul, choice, score) with no generated text. That is the same test the hosted d1 passed, with the license caveat above. The sibling d1-omni-600M is the smaller experimental checkpoint (text plus image, or text plus audio).