Running a language model locally with Ollama or LM Studio is great for privacy and cost, but every model is a multi-gigabyte download. Before you pull one, it helps to know whether it is any good for your use. Two new 1.7B-parameter models, small enough for a laptop or even a phone, came out this week: LFM3 1.7B from Liquid AI and Spark-X2.6 1.7B from XHToken. You can try both online first.
LFM3 1.7B (Liquid AI)
Liquid AI's new on-device model: fast, small, and good at chat, tool use and short reasoning. The family also includes vision (LFM3-VL), audio (LFM3-Audio) and omni (LFM3-Omni) variants of around 2B parameters, listed on the model page.
License: the LFM Open License 1.0, free to use for companies under US$10M in annual revenue.
LFM3 1.7BLiquid AI · LFM Open License 1.0 (free under $10M revenue)Liquid AI's 1.7B on-device model: fast, small and good at chat, tools and short reasoning.Try it online →Spark-X2.6 1.7B (XHToken)
A tiny chat model with controllable thinking: switch reasoning off for quick answers or on for step-by-step answers. There is also a larger 4B version. License: Apache-2.0, so commercial use is allowed without a revenue cap.
Spark-X2.6 1.7BXHToken · Apache-2.0A tiny 1.7B chat model with controllable thinking: small enough to run in a browser.Try it online →How fast are they?
On LiveWeights both run as the official 8-bit GGUF (Q8_0) builds with llama.cpp, the same format Ollama and LM Studio use. In our tests on a cloud GPU, LFM3 generated around 300 tokens per second and Spark-X2.6 around 240. On a laptop CPU they will be slower, but still usable, which is the point of models this size.
How to compare them
- Open LFM3 1.7B and Spark-X2.6 1.7B in two tabs.
- Ask both the same three or four questions from your real use (a summary, a rewrite, a small coding task).
- Try each with reasoning off and on: small models often improve a lot when they think first.
Expect mistakes: 1.7B models are fast and cheap, but they know less than large ones. If they are not enough, try a larger open model on the same site, such as Nanbeige4.3-33B or Apex 35B, before deciding what to download.
Then run it locally
Both models publish their GGUF weights on Hugging Face (linked on each model page), so once you have picked one you can load the same file in Ollama, LM Studio or llama.cpp on your own machine.