What is LFM2.5-350M?
LFM2.5-350M is one of the two smallest text models in Liquid AI's LFM2.5 family, a mix of convolution and attention layers designed to run on the device itself. Liquid trained it on 28T tokens, up from 10T for LFM2-350M, added large-scale reinforcement learning, and tuned it for data extraction, structured output and tool calls. It reads and writes text in 9 languages.
It is a current model. The newer LFM2.5-230M is a smaller sibling that scores lower on every test Liquid published, not a replacement. Liquid advises against using LFM2.5-350M for math, code, creative writing or questions that need broad knowledge.
What changed from LFM2-350M?
Same size, much more training: Liquid nearly tripled the pre-training data and added reinforcement learning, which lifted extraction, instruction following and tool use.
- Training
- Pre-training grew from 10T to 28T tokens, followed by large-scale reinforcement learning.
- Instruction following
- IFBench rose from 18.20 to 40.69, and IFEval from 64.96 to 76.96.
- Data extraction
- CaseReportBench rose from 11.67 to 32.45.
- Tool use
- BFCLv3 rose from 22.95 to 44.11, and BFCLv4 from 12.29 to 21.86.
Who is LFM2.5-350M best for?
LFM2.5-350M is best for high-volume data extraction, small on-device agents, structured output in 9 languages, and teams that want to fine-tune their own small model.
High-volume data extraction
Liquid measured up to 40.4K output tokens a second on a single H100, which suits batch jobs that pull fields out of millions of documents.
Small on-device agents
Its BFCLv3 tool-calling score rose to 44.11 from 22.95 on LFM2-350M, and LEAP, llama.cpp and MLX run it on phones and laptops.
Multilingual structured output
It handles 9 languages, including Arabic, Chinese, Japanese and Korean, and can return tool calls as JSON when the system prompt asks.
Teams training their own small model
A base checkpoint and ready-made fine-tuning notebooks are published, and the license lets you keep fine-tuned versions private.
What's included with LFM2.5-350M?
Native tool calling, base and instruct checkpoints, ready-made GGUF, MLX, ONNX and OpenVINO files, 4-bit QAD checkpoints, LEAP for phone apps, and published fine-tuning recipes.
- Native tool calling
- It writes function calls as a Python-style list between special tokens by default, or as JSON if you ask for it.
- Base and instruct checkpoints
- LFM2.5-350M-Base is for your own fine-tuning; the post-trained LFM2.5-350M is ready to use as it is.
- Ready-made formats
- Liquid ships GGUF, MLX, ONNX and OpenVINO versions on Hugging Face, with quick starts for Transformers, llama.cpp, vLLM and SGLang.
- 4-bit QAD checkpoints
- Quantization-aware Q4_0 files keep about 96.5% of the full-precision score while staying small and fast.
- Phone apps through LEAP
- Liquid's free LEAP platform packages the model into iOS and Android apps.
- Fine-tuning recipes
- Liquid publishes SFT, DPO and GRPO notebooks for Unsloth and TRL to adapt it to one task.
How much does LFM2.5-350M cost?
LFM2.5-350M is free to download, run and fine-tune, including in commercial products, for any company under $10 million in annual revenue; larger companies need a commercial license priced by Liquid AI's sales team.
- Download, run and fine-tune the open weights
- Commercial use allowed below the revenue limit
- Fine-tunes can stay private, with no copyleft
Companies with more than $10 million in annual revenue need Liquid AI's Enterprise commercial license, priced through sales by deployment scale and support needs. Research, education and qualified non-profit use is free with no revenue limit. There is no per-token charge: the model runs on your own hardware. The weights are on Hugging Face under the LFM Open License v1.0, and LEAP, Liquid's platform for deploying models on iOS and Android, is free to use. Full pricing on liquid.ai. We checked these prices on .
How do readers rate LFM2.5-350M?
No one has rated LFM2.5-350M yet. Be the first.
Sign in to rate. One rating per account.
LFM2.5-350M pros and cons
It is tiny, fast and far better than LFM2-350M at extraction and tool calls, but it knows little, reads only text, and its license caps free commercial use at $10M in revenue.
Pros
- Beats LFM2-350M by a wide margin on Liquid's tests: IFBench 40.69 against 18.20, CaseReportBench 32.45 against 11.67, and BFCLv3 44.11 against 22.95.
- Runs on modest hardware: Liquid measured 313 tokens a second on an AMD Ryzen AI CPU using 434MB, and partners ran it on an iPhone 13 Mini, a Pixel 6a and a Raspberry Pi 5.
- People who test small models on Hacker News and Reddit rate Liquid's among the best in their size class, and one found its 230M and 350M models far more usable for tool calls than FunctionGemma.
- Free for commercial use below $10M in yearly revenue, with no per-token bill and no duty to publish your fine-tunes.
Cons
- Not a general assistant: Liquid advises against math, code, creative writing and knowledge-heavy questions, and users note a 350M model holds little world knowledge, so pair it with search or fine-tune it.
- The license is not plain Apache 2.0: free commercial rights end once your company passes $10M in yearly revenue, and developers on Reddit and Hacker News have flagged that limit.
- Text only, with a 32,768-token context; images need a sibling such as LFM2.5-VL-450M, and Qwen3.5-0.8B reads 262,144 tokens.
- Multi-turn agent scores stay low in Liquid's own table (18.86 on τ²-Bench Telecom, 17.84 on Retail), so long agent loops need fine-tuning or a bigger model.
What are the best LFM2.5-350M alternatives?
Three tools on Recaply do LFM2.5-350M's job. Qwen3.5 Small and Gemma 4 are open source, so you can run them on your own server.
LFM 2.5FreeChoose it when 350M parameters are too few: Liquid recommends LFM2.5-1.2B-Instruct for most uses, and LFM2.5-2.6B adds a 128K context, all under the same free license.
Qwen3.5 SmallOpen SourceChoose it when you need image or video input, a 262K context or broader knowledge; Qwen3.5-0.8B ships under Apache 2.0 with no revenue limit.
Gemma 4Open SourceChoose it when you want audio and image input on a phone; its E2B size loads in about 1.1 GB on mobile, and Google's Gemini API serves it free.
Questions about LFM2.5-350M
What is LFM2.5-350M?
LFM2.5-350M is Liquid AI's 350-million-parameter open-weight text model for pulling data out of documents, writing structured output and calling tools on phones, laptops, cheap CPUs and busy GPU servers.
Who is LFM2.5-350M for?
LFM2.5-350M is best for high-volume data extraction, small on-device agents, structured output in 9 languages, and teams that want to fine-tune their own small model.
Is LFM2.5-350M free?
Yes. LFM2.5-350M has a free plan: download, run and fine-tune the open weights.
Is there an open source alternative to LFM2.5-350M?
Yes. Qwen3.5 Small and Gemma 4 are open source tools that do the same job, so you can run your own.
What are the best LFM2.5-350M alternatives?
The closest alternatives on Recaply are LFM 2.5, Qwen3.5 Small and Gemma 4.
Where can I try LFM2.5-350M?
On its official website, liquid.ai.
Can I use LFM2.5-350M in a commercial product?
Yes, as long as your company's annual revenue stays under $10 million. The LFM Open License allows commercial use and private fine-tunes; once revenue passes that line, the free commercial rights end and you need a paid license from Liquid AI.
Which tasks should I avoid with LFM2.5-350M?
Liquid AI does not recommend LFM2.5-350M for math, code, creative writing or knowledge-heavy questions. It is tuned for tool calling, data extraction and structured output, and at 350M parameters it trades broad knowledge for speed and a small footprint.
Can LFM2.5-350M run on a phone?
Yes. Liquid's LEAP platform deploys it in iOS and Android apps, and Liquid's published speed table shows it running on a Galaxy S25 Ultra, an iPhone 13 Mini and a Pixel 6a, each with well under 1 GB of memory.
Can LFM2.5-350M read images?
No. LFM2.5-350M takes in and returns text only. For pictures, Liquid offers separate vision models in the same family, such as LFM2.5-VL-450M, which you would run alongside it or instead of it.
Should I use LFM2.5-350M or LFM2.5-230M?
Pick LFM2.5-350M unless memory is very tight. In Liquid's published table it beats the newer LFM2.5-230M on every test, for example 32.45 against 22.51 on CaseReportBench and 18.86 against 5.26 on τ²-Bench Telecom.
Our verdict
LFM2.5-350M is a strong pick when you need a very small model to pull structured data or call tools on your own hardware.
It roughly doubles LFM2-350M's scores on Liquid's extraction, instruction and tool-calling tests, runs in well under 1 GB on phones and CPUs, and costs nothing below $10M in revenue.
- Skip it for
- Chat, code, math or anything that needs broad knowledge; a bigger LFM2.5 or Qwen3.5 model will serve you better.
- Check first
- Check your company's revenue against the $10M license line, and test it on your own documents first.
LFM3
Claude Haiku 5.5
GPT-6 Astra
Qwen3.7-Max
Amazon Nova
GPT-5.4