The liquid ai lfm 3 release landed on 27 October, and it is one of the more practical open-model launches for small teams: a 1.7B-parameter text model that Liquid AI says fits in about 1 GB of memory at 4-bit precision. This research-based post covers what was released, where it is strong and weak according to Liquid AI's own numbers, and how to try it locally without paying for an API.
What Liquid AI released
According to the Liquid AI LFM3 announcement, the first LFM3 release is a family of six open-weight models in the 2B range, all built on one shared backbone:
- LFM3-1.7B: a text model for local agents, with a 256K-token context window
- LFM3-VL-2B: vision, now including video input
- LFM3-Audio-2B: speech in and out, 18 languages
- LFM3-Omni-2B: vision, audio and text together
- d1-2B and d1-Omni-2B: "decision" models that answer a list of questions about a ticket, invoice, photo or voice note with probabilities, instead of writing free text
The weights are on Hugging Face under the LFM Open License v1.0. Read the license yourself before using it in a commercial product.
Where it is strong, and where it is not
The LFM3-1.7B model card says the model is designed for "local agents, tool calling, structured outputs, and domain-specific fine-tuning." In Liquid AI's published benchmark table it scores 71.39 on BFCL V4 (function calling) and 96.66 on IFEval (instruction following), ahead of the larger models it was compared with.
The same card is honest about the limits: it is "not recommended for agentic coding or complex multi-step reasoning due to its small size," and it suggests using retrieval or search for knowledge-heavy tasks. Treat these as vendor benchmarks until independent tests appear.
For a small business, that points to narrow, repetitive jobs:
- sorting incoming emails or support tickets into categories
- pulling fields out of invoices or order forms into JSON
- tidying files or tagging content in the background
- private tasks where you would rather not send data to a cloud API
A Liquid AI team member on Hacker News gave similar examples, such as classifying emails and cleaning a downloads folder.
How to try it locally
The model card lists ready-made checkpoints for different tools:
- GGUF for llama.cpp, Ollama and LM Studio (CPU or GPU)
- MLX 4-bit for Apple Silicon Macs
- ONNX for ONNX Runtime
- BF16 / FP8 for vLLM or SGLang on a GPU server
If you want a desktop app, LM Studio with the GGUF file is the simplest route. Check that your llama.cpp-based tool is a recent build, because the card notes that LFM3 needs LFM3-compatible llama.cpp builds.
If you prefer Python, the model card's quick start uses Transformers:
from transformers import pipeline
generator = pipeline(
"text-generation",
model="LiquidAI/LFM3-1.7B",
trust_remote_code=True,
)
messages = [{"role": "user", "content": "Sort this email: 'My order #1043 has not arrived.' Categories: billing, shipping, other. Reply with one word."}]
output = generator(messages, do_sample=True, temperature=0.1, top_k=50,
repetition_penalty=1.1, max_new_tokens=512)
print(output[0]["generated_text"][-1]["content"])
The card recommends temperature=0.1, top_k=50, repetition_penalty=1.1 for chat, tools and agents. Reasoning is on by default; you can pass enable_thinking=False to the chat template, though the card warns this can be inconsistent.
Practical tips before you rely on it
- Start with one narrow task and 20-30 real examples, and check its answers by hand.
- Ask for a fixed output format (one word, or JSON) so mistakes are easy to spot.
- Keep a human check on anything that sends money, emails customers or deletes files.
- Fine-tune only if prompting falls short. Liquid AI says every LFM3 model can be fine-tuned on a single GPU and links SFT, DPO and GRPO notebooks from the model card.
We post research-based AI news and how-tos for small businesses and creators at ainews.useaifix.com. If you want help setting up AI automation for your business, see useaifix.com.
Top comments (0)