Liquid LFM3: Six Small Open Models for On-Device Agents and Decisions

Question types (d1 model cards)

• • noul: yes or no; returns P(yes).
• • choice: one option from named criteria, with per-option probabilities; multiple: true makes it multi-label.
• • score: one of 2 to 10 ordered levels; returns the median level and its probabilities.
• • text, number, date: values read from the state, with their span; None when not found.
• • number with range: an estimate with a p10–p90 interval.
• • box and boxes: one object or every instance (up to 16) in a picture.

d1-2B decision benchmarks (self-reported)

Order: d1-2B / d1-Omni-2B / d1-3B (the October model built on LFM2.5), as published on the d1-2B card. Liquid scored all of them on the same items.

• • Decision Index public suite (v0.3): 51.1 / 49.0 / 49.0
• • DecisionBench (Hanno Labs) v1: 77.1 / 76.4 / 73.1
• • JevBench v1.4.1: 77.9 / 75.3 / 76.2
• • WorkflowEvals: 81.6 / 77.1 / 74.4
• • Typed Decisions: 61.3 / 63.1 / 65.0 (the older d1-3B wins this one)
• • Average of the public benchmarks every model on the card shares: 70.2 / 67.7 / 65.4
• • Latency for one question, warm, after model.compile(): 5.7 ms on an NVIDIA H100 PCIe, 3.8 ms on an AMD MI325X, 18.6 ms on an Apple M5 Pro

llama.cpp: the text model (LFM3-1.7B-GGUF card)

Needs b11960 or newer. The second command picks the quantization-aware distilled Q4_0 file, which Liquid trained to recover the accuracy lost to Q4_0.

# llama.cpp b11960 (v0.8.0) or newer
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M \
  --temp 0.1 --top-k 50 --repeat-penalty 1.1

# Quantization-aware distilled Q4_0 build (1.05 GB)
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF \
  --hf-file LFM3-1.7B-QAD-Q4_0.gguf \
  --temp 0.1 --top-k 50 --repeat-penalty 1.1

llama.cpp: d1-2B as a decision server (d1-2B-GGUF card)

A question is read in one batch, so -b and -ub must hold the state plus any media. In llama.cpp, d1 answers noul, choice and score questions.

llama-server -hf LiquidAI/d1-2B-GGUF:Q8_0 -b 8192 -ub 8192

curl http://127.0.0.1:8080/v1/systemone -H "Content-Type: application/json" -d '{
  "state": "I was charged twice this month, please refund one of them.",
  "questions": {
    "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},
    "team": {"type": "choice", "instructions": "Which team should handle this?",
             "criteria": {"billing": "Charges, refunds, invoices", "technical": "App or site faults",
                          "fraud": "Suspected unauthorised use"}},
    "urgency": {"type": "score", "instructions": "How urgent is this?",
                "criteria": ["Can wait", "Today", "Blocking the customer now"]}
  }
}'

llama.cpp: a question about a video (LFM3-VL-2B-GGUF card)

Video input needs ffmpeg in your PATH. For an interactive chat, run llama-cli with the same flags and type /image or /video to attach media.

llama-mtmd-cli -hf LiquidAI/LFM3-VL-2B-GGUF:Q4_0 \
  --video reading.mp4 -p "What is the child doing in this video?" \
  --temp 0.1 --top-p 1.0 --top-k 50 --min-p 0.0 --repeat-penalty 1.1

Ollama 0.40.12 on Apple hardware (release notes)

lfm3:latest is the 1.7B in NVFP4 on Ollama’s MLX engine (1.7 GB). Ask for d1:2b explicitly, because d1:latest still points to the older d1-3B.

ollama run lfm3

ollama run d1:2b "I was charged twice for order 4417." --ask "Is this a billing issue?"

Transformers: LFM3-1.7B (model card)

Loads Liquid’s own model code from the repo. Reasoning is on by default; see the sampling notes below.

# pip install -U transformers torch accelerate
from transformers import pipeline

generator = pipeline(
    "text-generation",
    model="LiquidAI/LFM3-1.7B",
    trust_remote_code=True,
)

messages = [{"role": "user", "content": "What is C. elegans?"}]
output = generator(
    messages,
    do_sample=True,
    temperature=0.1,
    top_k=50,
    repetition_penalty=1.1,
    max_new_tokens=2048,
)
print(output[0]["generated_text"][-1]["content"])

Transformers: d1-2B (model card)

One forward pass and no sampling parameters. The last line is the output printed on the card, not one I produced.

# pip install "transformers>=5.19" torch torchvision pillow
import torch
from transformers import AutoModel

device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
dtype = torch.float32 if device == "cpu" else torch.bfloat16
model = AutoModel.from_pretrained("LiquidAI/d1-2B", trust_remote_code=True, dtype=dtype).to(device)

# Several named questions over one state, answered in one pass
questions = {
    "refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"},  # yes/no: P(yes)
    "team": {"type": "choice", "instructions": "Which team should handle this?",
             "criteria": {"billing": "Charges, refunds, invoices", "shipping": "Delivery and parcels",
                          "technical": "App or site faults"}},
    "topics": {"type": "choice", "multiple": True, "instructions": "Which topics does the message raise?",
               "criteria": {"billing": "Payments and charges", "delivery": "Shipping and parcels",
                            "login": "Signing in to the account"}},
    "urgency": {"type": "score", "instructions": "How urgent is this ticket?", "criteria": ["low", "medium", "high"]},
}
out = model.system_one("I was charged twice and my parcel never arrived. Please refund one charge.", questions)
a = out["answers"]
urgency = a["urgency"]["legend"][str(int(a["urgency"]["score"]))]
print(round(a["refund"]["noul"], 2), a["team"]["choice"], a["topics"]["choices"], urgency)
# 1.0 billing ['billing', 'delivery'] medium

MLX: the 4-bit text model (once mlx-lm supports it)

This is the card’s command, and it won’t run on mlx-lm 0.32.2. The CLI has no repetition-penalty flag, so the card suggests make_logits_processors(repetition_penalty=1.1) in Python.

mlx_lm.generate --model LiquidAI/LFM3-1.7B-MLX-4bit \
  --prompt "What is the capital of Australia?" \
  --temp 0.1 --top-k 50 --max-tokens 16384