LFM3-1.7B: Local Setup, Memory and Model Choices

The post-trained LFM3-1.7B checkpoint accepts and generates text. Choose it for conversation, structured answers and tool use. The Base checkpoint is a pretraining starting point for adaptation; use the post-trained model for the first chat.

The family shares a backbone, but each download has different inputs and serving requirements. The decision models answer defined questions with structured predictions and probabilities, which calls for a different integration from a chat assistant.

Liquid describes the four-bit text model as roughly 1GB. Context cache and compute buffers add to that footprint; the operating system, browser and agent harness also need memory.

The model supports 256K tokens across input and output. Liquid reports a 1.6GB KV cache at 128K context for its text model. A short prompt and one request have a different budget from long documents or several simultaneous users. Start with a small context allocation and increase it with the workload.

The architecture combines short convolutions, sliding-window attention and global attention. Only part of the model keeps a full-length attention cache, which helps contain long-context memory. Measure the complete process on the intended runtime and quantization before committing device capacity.

Install or build llama.cpp with LFM3 support. Text and VL support was merged in upstream pull request 31252 on October 27, 2026. A client with an older embedded engine can reject the architecture even when it can download the GGUF. Update the inference engine as well as the model file.

The command below selects the official text GGUF at Q4_K_M, limits the initial context to 4,096 tokens and caps generation at 256 tokens. The first launch downloads the weights and needs internet access. The requested output is a simple checklist you can inspect before wiring the model into an application.

Pratik Gandhi published a launch-day test on an M5 Pro Mac mini with 64GB unified memory, macOS 27.0.1 and llama.cpp commit eb7818c using Metal, all layers on the GPU and flash attention. With Q4_K_M, llama-bench reported median generation rates of 184 tokens/second at 512-token context, 178.2 at 4K and 80.1 at 128K. Each point used one warmup and three timed runs with 256 generated tokens.

At 128K, llama.cpp allocated about 3.15GB on the GPU, including model, context and compute buffers. The three 128K generation runs ranged from 66.4 to 81.4 tokens/second. The reported 0.09-second and 0.72-second first-token figures at 512 and 4K were calculated from prompt throughput.

The same author’s small greedy-decoding smoke test also shows why decode speed is only part of the decision: with thinking enabled, a request for an is_prime function was still reasoning at its 1,024-token cap and produced no final answer.

The checkpoint can generate tool calls, but your application supplies the schemas, parses the call, executes it and returns the result. The LFM3 chat template emits XML-style function calls. A parser built for an older Liquid model needs to match this format before it can reliably run an agent loop.

Pass tool schemas through the chat template and retain the current tool-use turn’s content and reasoning when returning results. Reasoning is enabled by default; the model card says disabling it can be inconsistent on reasoning-heavy tasks. Check the saved result of a small workflow before allowing longer sequences.

Liquid positions this model for narrow agents and specialization. Its model card advises retrieval or search for knowledge-heavy tasks and recommends larger models for complex reasoning or agentic coding.

A local run uses your device’s electricity, memory and time. For a rented machine, multiply the full reserved time by the hourly rate, including model loading and retries. For example, an assumed $0.50/hour machine reserved for 20 minutes costs about $0.167 before storage and transfer.

Liquid also offers a hosted Playground and API. Those are separate usage routes with their own account and billing conditions. Choose local inference when device control and offline operation matter; use the hosted entry to explore the model before downloading a runtime.

The weights use the LFM Open License v1.0. Its commercial-use section includes an annual-revenue threshold defined as $10 million and limits the grant for entities exceeding that threshold. Read that condition before adopting the weights for a business.