Model guide
LFM3 1.7B
LFM3 1.7B is a small text model for extracting fields, choosing tools and specialized local tasks. Start with a short request whose answer you can check, including a request with missing information. Its 256K context limit does not mean that a laptop can load that much text, and its roughly 1.1 GB Q4 download is only the model weights.
At a glance
- Parameters
- 1.73B
- Accepts
- Text
- Returns
- Text
- Published context limit
- 262,144 tokensInput and output together. The fit calculation conservatively counts all 11 attention layers at the selected context length; a runner that retains only 256 tokens in the five sliding-window layers can use less conversation memory.
- Model format
- GGUF · safetensors · ONNX
- Software and download hosts
- MLX. Follow the exact setup below; this list does not establish support for every format, input type or software version.
- Suggested compression
- Q4_K_M
- Model-weight download
- Q4_K_M weights · 1.10 GBLiquidAI (model publisher)LFM3-1.7B-Q4_K_M.gguf
- Languages
- English · Arabic · Chinese · French · German · Italian · Japanese · Korean · Portuguese · Spanish · Vietnamese · Thai · Indonesian · Hindi · Russian · Polish
- License
- LFM Open License v1.0 (commercial use is subject to the US$10 million annual-revenue threshold in section 5)
Will it fit?
Choose the computer and memory you plan to use. Not sure keeps the answer incomplete; these controls do not detect your hardware.
Enter the positive amount your computer shows. Leave unknown memory as Not sure.
LFM3 1.7B · Q4_K_M · 4,096 tokens · one session · llama.cpp estimate
Enter the missing information before choosing a model.
Choose a device and memory type. Enter system RAM or unified memory. Choose a graphics processor, or choose no dedicated GPU if the model will use system RAM.
RAM: not supplied. Estimated need: 3.0 GiB system RAM.
Memory type has not been supplied; choose it before relying on this estimate.
This estimate uses the Q4_K_M model-weight download size, llama.cpp, 4,096 tokens of text context, FP16 conversation memory, one user and a 20% safety allowance. Conversation memory uses the published layer layout; app memory and temporary workspace are allowances. Fit is checked after the operating-system or display reserve. These are estimates, not measured use, and do not establish GPU or architecture support. Image/audio processing, companion files and document indexes need additional memory.
Best for
- Extract dates and quantities from short requests. Try a second request that omits both fields and check that it returns null for each.
Choose something else when
- If you need an agent to act, test the proposed tool name and arguments separately before allowing execution.
First setup
Check a representative task
Open Model-specific setup and copy the delivery-request test. In Ollama, run it twice, once with a date and count and once with both missing. Expect two nulls on the second run. Valid JSON alone does not prove that the fields are correct or that a runner enforces structured output.
Extract a delivery request
Use 4,096 context tokens and a 256-token reply allowance. Follow Liquid AI's general settings: temperature 0.1, top-k 50 and repetition penalty 1.1 where your runner exposes them. Keep reasoning enabled initially; it can add waiting time. This exercise extracts text and does not connect or run tools.
Use only this request. Return a JSON object with exactly these keys:
- delivery_date: the date exactly as written, or null if missing
- item_count: an integer, or null if missing
Do not invent missing information or add commentary.
Request: Deliver 3 notebooks on 2026-11-12. Leave them at reception.
Check for exactly {"delivery_date":"2026-11-12","item_count":3} after any separately displayed reasoning. Parse the JSON and check its values, not just its syntax. Start a new conversation with the same instructions but change the request to Leave the notebooks at reception. Both values must be null. An invented date or count fails the check.
Then test representative requests you have labelled yourself. This short exercise does not establish reliable extraction, native tool execution or enforced structured output. Use document or web search for knowledge-heavy questions, and a larger model for difficult coding or multi-step reasoning.
In LM Studio's Discover tab, search for LiquidAI/LFM3-1.7B-GGUF and download LFM3-1.7B-Q4_K_M.gguf (about 1.10 GB). Update the llama.cpp runtime through LM Studio's runtime manager if LFM3 is not supported. In Chat, select this text model and set context to 4,096 before loading. Use the extraction and missing-field checks and their sampling settings. If loading fails with an unsupported-architecture error, update the runtime; if allocation fails, shorten the context or free memory. Stop generation with Stop, then unload from the model selector when finished.
Install a current llama.cpp release with LFM3 support. In its executable folder run:
./llama-server -hf LiquidAI/LFM3-1.7B-GGUF -hff LFM3-1.7B-Q4_K_M.gguf -c 4096 --host 127.0.0.1 --port 8080
On Windows PowerShell use .\llama-server.exe in place of ./llama-server. Wait for the download and model load, then open http://127.0.0.1:8080. Use the extraction and missing-field checks, a 256-token reply allowance and the sampling settings in that example. An unsupported-architecture error needs a newer runtime; an allocation failure needs smaller settings or more memory. Press Ctrl+C in the server terminal to stop it.
Load the model, start a fresh chat and time your usual prompt. Record time until the first useful word and time until the full answer. Note the processor, context and output length when comparing another run.
- Redistribution requires the license and retained copyright and attribution notices; read the controlling license.
- LFM3 1.7B keeps only the last 256 tokens on 5 of its 24 layers. The estimates on this page count conversation memory (KV cache) as if those layers stored the whole conversation, the cautious choice (0.09 GiB at 4,096 tokens). A runner that caps those layers needs less, about 0.05 GiB.
- The roughly 1.10 GB Q4 file is model weights, not total memory. The runner, conversation memory, operating system and safety allowance also need room.
- Six full-attention and five sliding-attention layers store conversation memory. Thirteen convolution layers retain fixed-size state. The estimate conservatively counts sliding attention at full context and uses FP32 convolution state.
- The text model does not accept images or audio. Liquid AI advises using search for knowledge-heavy tasks and does not recommend it for agentic coding or complex multi-step reasoning.