Can You Run LFM3-1.7B Locally?

Yes, on almost any GPU. LFM3-1.7B is a 1.73-billion-parameter text model that Liquid AI published on Hugging Face on October 23, 2026, with an official GGUF repository the next day. Computed from the real GGUF files, the Q4_K_M build needs 1.87 GB at a 4,096-token context, so even a 4 GB card runs it fully on the GPU. The license is the LFM Open License v1.0, which caps commercial use by company revenue.

The short answer by GPU

Memory is the real size of the GGUF file plus the KV cache at a 4,096-token context and a fixed 0.8 GB system margin. Your desktop and browser also use some VRAM, so on a 4 GB card that runs a display, Q6_K (2.18 GB in total) leaves more room than Q8_0.

What "about 1 GB" means in practice

Liquid AI says the model takes about 1 GB at 4-bit precision. That matches the file: the Q4_K_M GGUF is 1.02 GB. What the vendor figure leaves out is everything around the weights. At a 4,096-token context the total is 1.87 GB, because 0.05 GB of KV cache and the 0.8 GB margin come on top. Plan your VRAM around the total, not the file size.

The repository also has a Q4_0 file built with what Liquid AI calls Quantization-Aware Distillation: a checkpoint trained to win back the accuracy that Q4_0 quantization costs. It is the smallest file in the repository at 0.98 GB (1.82 GB total at 4K). We have not compared its output quality with Q4_K_M; that claim is the vendor's.

Memory per quantization

We read the file list of the official GGUF repository on October 31, 2026 (129,608 downloads, last modified October 28). Each quantization is a single file. The KV cache column assumes the default 16-bit cache.

Architecture notes: why long context stays cheap

The config.json lists 24 layers, and only 6 of them are global attention layers that keep a full KV cache. Thirteen are short convolutions with a tiny fixed state, and five use sliding-window attention over the last 256 tokens. Liquid AI calls this mix "tri-hybrid".

For your memory budget, this is the part that matters. At 32,768 tokens the six global layers need 0.38 GB of KV cache. If all 24 layers kept a full cache, the same context would need 1.50 GB, four times more. At the native 256K context, the cache reaches 3.00 GB, which is still small enough for Q4_K_M to stay under 6 GB in total. The sliding-window layers add only a few megabytes, which the calculator leaves out.

How to run it

ollama run lfm3
ollama run hf.co/LiquidAI/LFM3-1.7B-GGUF:Q8_0
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q8_0 -ngl 99 --temp 0.1 --top-k 50 --repeat-penalty 1.1

Check your runtime version first. Ollama added LFM3 in version 0.40.12 for Apple hardware only; version 0.40.13, released October 28, extended it to Windows and Linux. On llama.cpp you need build b11960 (v0.8.0) or newer, and LM Studio needs its llama.cpp engine 2.65.1 or newer, according to the GGUF repository. The sampling settings above are the ones the model card recommends for general use.

Reasoning is on by default. The model card says you can turn it off with enable_thinking=False, but warns the switch can behave inconsistently.

Which one should you pick?

On a 4 GB card, use Q8_0 if the GPU is free, or Q6_K if it also drives your display. On 6 GB, BF16 fits at 4K and 32K contexts; for the full 256K window, drop to Q8_0 (5.52 GB), or to Q4_K_M (4.82 GB) if the card also drives your display. On 8 GB and above, BF16 fits even at 256K (7.03 GB). Liquid AI positions the model for local agents, tool calling and structured output, and reports a BFCL V4 tool-use score of 71.39 against 62.83 for Qwen3.5 9B; those are the vendor's numbers, not ours. The model card itself advises against using it for agentic coding or complex multi-step reasoning.

Limits

• License: the LFM Open License v1.0 grants commercial rights only to users whose organization stays below 10 million US dollars in annual revenue. Read the license text before any commercial use. The repository is not gated.
• No speed figure here. The editor has not run LFM3 yet, and our measurement on an 8 GB laptop GPU showed real speeds between 23% and 81% of the calculator's estimate, so an estimate alone would say little.
• The other LFM3 models (vision, audio, omni, decision) ship extra encoder files; this guide covers the text model only.

See every quantization live on the LFM3-1.7B model page, check it against your own GPU, or compare with other small models you can run anywhere. Sources: the official model card and Liquid AI's LFM3 announcement.

How we made this: file sizes and architecture were read from the public Hugging Face API on October 31, 2026; memory totals were computed with the AI Local Check calculator (see our methodology); the draft was written with AI assistance and reviewed by the editor before publication.