LFM2.5-230M
LFM2.5-230M is Liquid AI's smallest LFM2.5 model: a 230M-parameter, text-only hybrid (14 layers — 8 double-gated LIV convolution blocks plus 6 GQA attention blocks) trained on a 19T-token budget and distilled from LFM2.5-350M with multi-stage reinforcement learning. Context is 32,768 tokens; it covers ten languages including English, Chinese, Japanese, and Spanish. Vendor-measured speeds: 213 tok/s decode on a Galaxy S25 Ultra and 42 tok/s on a Raspberry Pi 5. It ships in native, GGUF, ONNX, and MLX 8-bit formats with llama.cpp, LM Studio, vLLM, and MLX support, under the LFM Open License v1.0 — free, including commercial use, below $10M annual revenue. Liquid pitches it for tool use and data extraction; reported benchmarks include IFEval 71.71, BFCLv3 43.26, and GPQA Diamond 25.41.
Positioning
This is the rare sub-1B model with an actual job. Run it if you're embedding an always-on model in a phone app, a Raspberry Pi project, or an on-device agent pipeline where the model routes function calls and extracts structured data rather than holding a conversation. The vendor-measured 213 tok/s on a Galaxy S25 Ultra and 42 tok/s on a Pi 5 are the point: real-time inference on hardware where a 7B model is a slideshow. On instruction following it punches up — IFEval 71.71 beats Gemma 3 1B IT (63.49) and both Granite 4.0 350M variants — and BFCLv3 43.26 is strong for the size class.
Skip it if you want general intelligence: GPQA Diamond 25.41 and MMLU-Pro 20.25 mean knowledge depth is near zero, and Liquid itself warns against math, code generation, and creative writing. Qwen3.5-0.8B beats it on knowledge tasks if you can afford 3–4x the footprint.
The license needs eyes-open reading: LFM Open License v1.0 is free — commercial use included — only for entities under $10M annual revenue. Above that threshold, commercial use simply isn't licensed and you need a Liquid agreement, so treat it as source-available rather than open if you're a larger org. Hobbyists and startups are unaffected. Fine-tuned to one narrow task, it's about the best capability-per-watt available at this size.
Overview
LFM2.5-230M is Liquid AI's smallest LFM2.5 model: a 230M-parameter, text-only hybrid (14 layers — 8 double-gated LIV convolution blocks plus 6 GQA attention blocks) trained on a 19T-token budget and distilled from LFM2.5-350M with multi-stage reinforcement learning. Context is 32,768 tokens; it covers ten languages including English, Chinese, Japanese, and Spanish. Vendor-measured speeds: 213 tok/s decode on a Galaxy S25 Ultra and 42 tok/s on a Raspberry Pi 5. It ships in native, GGUF, ONNX, and MLX 8-bit formats with llama.cpp, LM Studio, vLLM, and MLX support, under the LFM Open License v1.0 — free, including commercial use, below $10M annual revenue. Liquid pitches it for tool use and data extraction; reported benchmarks include IFEval 71.71, BFCLv3 43.26, and GPQA Diamond 25.41.
Family & lineage
How this model relates to others in its lineage. Family members share architecture and training-data roots; parent / children edges record direct distillation or fine-tune relationships.
Strengths
Weaknesses
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of LFM2.5-230M.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
Can I use LFM2.5-230M commercially?
What's the context length of LFM2.5-230M?
Source: huggingface.co/LiquidAI/LFM2.5-230M
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify LFM2.5-230M runs on your specific hardware before committing money.