Liquid AI LFM3 1.7B packs a 256K context and native tool calling into about 1 GB at 4 bit precision. This tutorial runs it locally with Ollama and with llama.cpp, picks the right file for your memory budget, and sets the context size on purpose instead of by accident.
A model that fits in about 1 GB changes where the work can happen. It can live on the phone in your pocket, on the laptop you already own, and on a small server that never talks to a cloud API. Liquid AI built LFM3 1.7B for exactly that spot. This tutorial gets the text model running on your own hardware, first through Ollama for the shortest route, then through llama.cpp for full control over the file and the context.
What You Will Build or Learn
You will end with LFM3 1.7B answering prompts on your own machine, with no account and no cloud call in the loop. You will know which file size you picked and why, what context length you actually loaded, and how to serve the model locally for other apps to use.
The model itself is a 1.73 billion parameter text model from Liquid AI, released on October 27, 2026 as part of a six model family. Liquid AI lists it for local agents, tool calling, structured outputs, and work in a chosen domain after tuning. The company also draws a clear line in its own docs. Use search or retrieval for work that needs broad knowledge, and do not pick this model for agentic coding or for complex reasoning that needs many steps. Small size is the point, and small size is also the limit.
Before You Begin
Read the two catches first, because both are easy to miss in the headline numbers.
The first catch is memory. About 1 GB is the size of the model weights at 4 bit precision. It is not the total memory of a running session. The runtime needs memory, and the context needs memory, and that second bill grows as you load more tokens. A 256K context is a maximum the model supports, not a setting you should load by default. Start small, confirm the model works, and raise the context only for work that needs it.
The file is small. The context is not. House illustration by Dirty Static.
The second catch is the licence. LFM3 ships under the LFM Open License v1.0, which Liquid AI bases on Apache 2.0 with one key change. Free commercial use ends when a company passes 10 million US dollars in annual revenue, and larger companies must contact Liquid AI for a commercial licence. Personal use and research sit well inside the free terms. Product plans need a licence check before they ship.
Step 1: Pick a Runtime and Check Its Version
Action: choose one route and confirm the version is new enough. All three routes below come from Liquid AI and its model cards.
| Route | Best for | Version to check |
|---|---|---|
| Ollama | The shortest route to a first answer | Version 0.40.13 or newer on Linux and Windows, and 0.40.12 or newer on macOS |
| llama.cpp | Full control over the file, the context, and the local server | Build b11960, sold as version 0.8.0, or newer |
| LM Studio | A desktop app instead of a terminal | LM Studio with llama.cpp engine 2.65.1 or newer |
ollama --version
llama-cli --version
Expected result: at least one command prints a version at or above the line in the table. If both print older numbers, update the runtime first. On a Mac, the common installs are brew install llama.cpp and the Ollama app from its official site. On Windows, winget install llama.cpp is the common llama.cpp route. On Linux, build llama.cpp from its project repo or use a current release build, then check the version again.
Step 2: Run It With Ollama, the Short Route
Action: pull the model from the Ollama library and start it. The library name is lfm3, and the size tag names the 1.7B text model in full.
ollama pull lfm3
ollama run lfm3
You can also name the size in the command:
ollama run lfm3:1.7b
Expected result: Ollama downloads the model, then opens a prompt. Ask it a plain question with a checkable answer, such as asking it to sort a short list you give it or to turn a sentence you paste into a JSON object with named fields. A reply in the terminal means the model is loaded and working on your machine. Type /bye to leave the session.
Note what this route hides. Ollama manages the file and the defaults for you, which is why it is fast to start. The price is control. When you need a chosen file, a chosen context size, or a local server with set sampling values, use the llama.cpp route in the next steps instead. You can run both routes on one machine. They are separate installs and they do not share model files.
Step 3: Pick the Right GGUF File for llama.cpp
Action: choose a file size before you download anything. Liquid AI publishes the GGUF build as LiquidAI/LFM3-1.7B-GGUF on Hugging Face, and the docs name Q4_K_M as the balance point for size and quality in the GGUF set.
| File | Size on disk | Pick it when |
|---|---|---|
| QAD Q4_0 | 1.05 GB | You want the smallest file, and you want the build Liquid AI trained to recover quality lost to 4 bit work |
| Q4_K_M | 1.1 GB | You want the standard balance of size and quality for daily local use |
| Q8_0 | 1.85 GB | You have memory to spare and want higher precision |
| BF16 | 3.47 GB | You want the full precision file for tuning or for GPU serving work |
Expected result: you know the tag you will load in Step 4. If you are unsure, pick Q4_K_M. It is the file the rest of this tutorial uses, and you can swap the tag later without changing any other step.
Step 4: Run It With llama.cpp
Action: start the model in the terminal with the sampling values Liquid AI lists on the model card. The -hf flag downloads the file from Hugging Face on the first run.
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M --temp 0.1 --top-k 50 --repeat-penalty 1.1
For the QAD file from Step 3, name the file in full instead of using a tag:
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
Expected result: llama.cpp downloads the file, loads it, and opens its chat prompt. The load lines show the model size and the context it set aside. Ask the same checkable question you used in Step 2. For maths and science work, Liquid AI lists a second set of values. Swap in --temp 1.0 --top-p 0.95 --top-k 20 and drop the repetition penalty for that work, per the model card.
Reasoning is on by default in this model, per the Liquid AI docs. That means replies can include thinking before the final answer. It helps on tasks with steps. It also costs tokens, so judge the model on the final answer, not on the speed of the first word.
Step 5: Set the Context on Purpose and Serve It Locally
Action: start the local server with a context size you chose. The -c flag sets the context in tokens. Start at 8192, which covers most chat and agent work while keeping memory in check.
llama-server -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M --host 127.0.0.1 --port 8080 -c 8192 --temp 0.1 --top-k 50 --repeat-penalty 1.1
Serve it on the local address first. House illustration by Dirty Static.
Expected result: the server starts on your own machine at port 8080 and prints that it is listening. The host value in the command keeps it on the local address only. That is the right default. A local model server has no sign in of its own, so do not change the host to a public address unless you add protection in front of it first.
Raise the context only when the work needs it. A document task that loads a long file may earn -c 32768 or more. Watch memory as you raise it, because the context store grows with the setting even before you paste a long prompt. The 256K figure is the ceiling the model supports. It is not a target for a laptop session.
Verification
Prove the build works three ways, in this order.
First, check the server health from a second terminal while the server from Step 5 runs:
curl http://127.0.0.1:8080/health
A healthy reply means the server loaded the model and is ready. Next, send it one request through its chat address and read the reply in full. Pick a prompt with a right answer you can check by eye, such as a sum, a sort, or a short extraction from a paragraph you paste. Last, cut the network. Turn on airplane mode or unplug the cable, then repeat the prompt in Ollama or in llama.cpp. The model should answer in full with no network at all. That last test is the whole point of the build, so do not skip it.
If you used only the Ollama route, run ollama list and confirm lfm3 sits on disk, then repeat the offline test inside ollama run lfm3. If you used only llama.cpp, the health check and the offline test together prove the build.
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Unknown model architecture lfm3 | The runtime predates LFM3 support | Update to llama.cpp b11960 or newer, or to Ollama 0.40.13 or newer on Linux and Windows, then run the same command again |
| The download works but the load fails at once | The file and the runtime do not match, or the chat template is stale | Load a GGUF file with the current chat template, per the Liquid AI docs, or pass a current template file with the template flag the docs name |
| The machine runs out of memory on load | The context setting is too large for the machine | Restart the server with a smaller context, starting at -c 8192, and raise it only as the work needs |
| Answers drift or repeat on agent tasks | The sampling values do not match the card, or the task needs many reasoning steps | Set temperature 0.1, top k 50, and repetition penalty 1.1 for general work, and pick a larger model for complex work with many steps |
| The model invents facts about the world | The task needs broad knowledge, and the model card warns against that use on its own | Add search or retrieval for facts, and use the model for tool calls, structure, and local work |
| Ollama on Linux or Windows says the model needs a newer version | The install predates the LFM3 support release for that system | Update Ollama to 0.40.13 or newer, check the version, and pull the model again |
Next Steps
Put the local server to work. Liquid AI lists Hermes Agent, Pi, and OpenClaw as the harnesses this model was tuned for, so a local agent setup is the natural next build, and our Hermes Agent setup guide covers that install in full. Keep the server on the local address while you learn, and give the agent a test folder before you give it real files.
For document work, pair the model with retrieval instead of loading whole archives into the context. The long context helps with one long document. Retrieval helps with many documents, and it keeps memory low. If the work grows past this model, the same two routes in this tutorial carry the larger LFM builds and other GGUF models with no change in shape. Only the file tag and the memory bill change.
Commands, file sizes, and version lines checked against the Liquid AI LFM3 1.7B documentation, the Liquid AI model library notes, and the LFM3 1.7B GGUF model card, November 2026. This procedure was verified against those documents. Run it on your own machine and use the Verification steps above to confirm your install.
Sources
- Model documentation: Liquid AI LFM3 1.7B docs
- Model library and format notes: Liquid AI Liquid Foundation Models
- GGUF files and runtime version lines: LiquidAI LFM3 1.7B GGUF on Hugging Face
- Licence terms: Liquid AI LFM License
- Launch post: Liquid AI, Introducing LFM3, October 27, 2026
- House illustrations: Dirty Static
END OF TRANSMISSION
Stay curious. Question everything.
