Liquid AI released LFM3 on 27 October 2026: a family of small open-weight models whose text model, LFM3-1.7B, has a 1.10 GB Q4 download and a published limit of 256K tokens of context. Total memory also includes conversation memory and the runner; check the calculator with your hardware before loading it. Ollama added the family the same day. It is a candidate for sorting, extracting and calling tools on modest hardware. Liquid's model card advises using search for knowledge-heavy questions and a larger model for difficult coding or multi-step reasoning.
What Liquid released
LFM3 is six models of roughly 2 billion parameters ("2B") each, all built from one shared base model, LFM3-1.7B-Base:
| Model | Input types | What it is for |
|---|---|---|
| LFM3-1.7B | Text | Chat, tool calls, structured answers, fine-tuning |
Vision model (lfm3-vl in Ollama) | Text and images | Reading screenshots, photos and diagrams |
Audio model (lfm3-audio) | Text and speech | Speech tasks |
Omni model (lfm3-omni) | Text, images and speech | Combined tasks |
Two decision models (including d1 in Ollama) | Specialized from the vision and omni models | Picking a label or option instead of writing prose |
The text model has 1.73 billion parameters and a published context limit of 256K tokens, covering input and output together. It handles 16 languages, including English, Spanish, French, German, Chinese, Japanese, Hindi and Arabic. Liquid also publishes the base models for fine-tuning on a single GPU.
How much memory it needs
This is the headline for most readers. Liquid says the text model "fits in about 1 GB at 4-bit precision", and Ollama lists its downloads at 1.1 to 1.7 GB depending on the model compression (quantization).
The download is only part of the memory you need. Long conversations and documents also use conversation memory (KV cache), and that is where LFM3 stands out. Liquid reports 1.6 GB of conversation memory at a full 128K-token context for LFM3-1.7B, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. A model that is small on disk can still exhaust memory once you paste in a long document; LFM3 is built to avoid that.
The official Q4_K_M file is 1,099,627,808 bytes (about 1.10 GB); that is the model download, not total memory. Check your own computer with LFM3 selected in the RAM and VRAM calculator. It starts at 4,096 context tokens and leaves your hardware unknown until you enter it. You can also read how much RAM and VRAM local models need and the 8 GB RAM setup guide.
What the benchmarks say, and what they don't
All figures below are Liquid's own, compared with its previous small model, LFM2.5-2.6B:
| Benchmark | What it measures | LFM3-1.7B | LFM2.5-2.6B |
|---|---|---|---|
| BFCL V4 | Calling tools and functions correctly | 71.39 | 55.89 |
| WebArena-Verified | Completing tasks on websites | 38.42 | 3.02 |
| IFEval | Following formatting instructions | 96.66 | 95.39 |
| RULER 128K | Using information deep in a long context | 65.56 | 1.15 |
| GPQA Diamond | Graduate-level science questions | 63.37 | 57.78 |
The tool-calling and long-context scores are large jumps for a model this size. Liquid says LFM3 also beats the much larger Qwen3.5-9B on tool calling (71.39 against 62.83) and web tasks (38.42 against 28.94).
On Liquid's GPQA Diamond comparison, Qwen3.5-9B scored 80.40 and LFM3 1.7B scored 63.37. That test measures answers to difficult science questions. For a delivery-note extractor, measure missing fields and invented values on your own notes. Liquid's model card advises using document search or web search for knowledge-heavy tasks, and says the model is "not recommended for agentic coding or complex multi-step reasoning." Read the coding benchmark in Liquid's announcement with that warning in mind.
Run LFM3 in Ollama
You need Ollama 0.40.13 or newer on Windows or Linux. Ollama 0.40.12 added LFM3 but ran it only on Apple hardware; 0.40.13, released on 28 October, extended it to Windows and Linux. If you do not have Ollama yet, follow the Ollama installation guide.
ollama --version
ollama run lfm3
The first command confirms your version. The second downloads the text model (about 1.1 to 1.7 GB) and opens a chat. The download needs an internet connection; after that, the model runs on your computer.
Liquid recommends a low temperature for chat, tool calls and structured output. Inside the chat, enter:
/set parameter temperature 0.1
/set parameter top_k 50
/set parameter repeat_penalty 1.1
For math and science questions, Liquid recommends temperature 1.0, top_p 0.95 and top_k 20 instead.
The model reasons before answering by default. That usually improves answers but adds waiting time. Liquid notes that turning reasoning off can make answers less consistent on reasoning-heavy tasks, so leave it on unless speed matters more.
LM Studio and llama.cpp can run the GGUF model files that Liquid publishes, and Apple silicon Macs can use a 4-bit MLX version. Not sure which to use? See GGUF vs MLX.
Where LFM3 is worth trying
- Sorting and extraction. Pulling dates, names and amounts out of emails or forms into JSON.
- Tool routing. Deciding which function or app a request should go to, as the small "brain" of a local agent.
- Long inputs on small machines. Summarizing or searching a long transcript on a laptop that cannot hold a 9B model with a long context.
- Fine-tuning. Training a narrow specialist on your own examples on one GPU.
A quick test before you rely on it: take 20 real examples from your own work, run them through LFM3 and a larger model you already use, and count the wrong or missing answers in each. Separate a missing value from an invented one. For each model, count how often it fills in a delivery date your example omits and how often it misses a date that is present. Decide which errors require human review.
The license has a revenue limit
LFM3 uses Liquid's LFM Open License v1.0, the same license as LFM2.5. Non-commercial and research use is allowed with no revenue limit, and anyone who shares the model or a modified version must pass on the license and keep Liquid's copyright and attribution notices. Commercial use is licensed only while the user's organization, counted together with its affiliates, stays below $10 million in annual revenue. The license does not cover commercial use by a larger organization, which needs a separate agreement with Liquid. Qualifying non-profits are exempt from the limit for non-commercial and research use.
Converting the model to GGUF or running it through Ollama does not change those terms. If your employer is above the limit, check with Liquid before using LFM3 for work.
Keep it local
ollama run lfm3 runs the model on your own computer. Liquid also offers the models through its hosted API and online playground; text you type there goes to Liquid's servers. If an app offers "LFM3" without saying where it runs, check before sending private material. The local AI privacy guide explains what to look for.
Related reading
- Our earlier coverage of Liquid AI's LFM2.5 models
- What local LLM can my computer run?
- Local AI without a dedicated GPU
- Context windows explained