Liquid AI's New On-Device Model Family Is "Open" Only Below $10 Million in Revenue
Liquid AI released LFM3 today, a family of six open-weight models built around a shared 1.7-billion-parameter backbone trained on 22 trillion tokens. The lineup spans a text model, a vision model that accepts video, a speech-in-speech-out audio model, a combined omni model, and two "decision" variants that return calibrated probabilities instead of generated text. All six sit in the roughly 2-billion-parameter class, aimed squarely at phones, laptops, and edge hardware rather than data-center clusters.
The release landed with a wide hardware and distribution net: Ambarella, AMD, Arm, Intel, Nvidia, OpenRouter, Qualcomm, and Vercel are listed as day-zero partners, with Mercedes-Benz and Shopify named as pre-release testers. The weights are available on Hugging Face, Liquid's own Playground and API, with native support for llama.cpp, MLX, and vLLM; Ollama, LM Studio, and Unsloth are credited as ecosystem partners. On Ollama's library, four brand-new listings — lfm3, lfm3-omni, lfm3-vl, and lfm3-audio — appeared with zero recorded pulls, consistent with a launch that was only hours old at the time of writing.
Free, until it isn't
The models ship under the LFM Open License v1.0, not Apache or MIT. Its terms, laid out on Liquid AI's own legal page, grant unlimited free use with no user-count cap right up until an organization's annual revenue hits $10 million. Past that line, commercial use "is not licensed under this Agreement" at all — a company has to contact Liquid's sales team for a separate commercial license, and the license terminates automatically if its terms are violated. Qualified nonprofits (entities exempt under 501(c)(3) or a foreign equivalent) using the model for non-commercial or research purposes are exempt from the revenue cap — a narrower carve-out than "research use" generally, since a for-profit company's own research use would not qualify. There's no copyleft requirement, so fine-tuned derivatives don't have to be shared back.
That structure puts LFM3 in the same category as a growing list of "open-weight, not open source" releases from frontier and near-frontier labs this year: free enough to seed a developer community and show up in benchmarks, restrictive enough to funnel any company that actually succeeds with the model back into a paid contract. The $10 million threshold is generous compared to some peers' limits, but it's a hard line, not a sliding scale — and it means LFM3 doesn't meet the Open Source Initiative's definition of open source regardless of how freely the weights circulate.
The benchmarks are Liquid's own
Liquid's launch materials claim LFM3-1.7B beats its predecessor, LFM2.5-2.6B, using 35% fewer parameters, and edges out Qwen3.5-9B on an average score (49.37 versus 47.24) despite being roughly a fifth the size. The vision model, LFM3-VL-2B, is pitched as matching Gemma 4 12B on a 62-benchmark average (71.4 versus 71.7) at a sixth of the parameter count, and the omni model claims wins on Daily-Omni, WorldSense, and AV-SpeakerBench against models up to 10.7 billion parameters, though it trails the larger Qwen2.5-Omni-7B on the broader OmniBench and RealWorldQA benchmarks. Independent coverage from Unite.AI and AlphaSignal both note that every one of these figures comes from Liquid's own release materials; neither found a third-party lab that had reproduced them by publication time.
AlphaSignal's side-by-side read of the same release materials also surfaced the caveats Liquid's framing tends to skip past: agentic vision performance lags well behind frontier systems (a 42.5 average, described as a "substantial gap"), general audio understanding trails the smaller MiMo-Audio-7B, and the headline latency numbers exclude real-world overhead like network transit and endpoint detection. A wind-turbine inspection demo comparing LFM3 favorably to a GPT model wasn't a like-for-like test, either — only the LFM3 setup had received task-specific training for that job.
The base model nobody has actually benchmarked yet
The quieter release in the family is LFM3-1.7B-Base, the raw pretrained checkpoint that developers would actually fine-tune for their own use. An independent look at its Hugging Face listing found that Liquid published no benchmark table for it at all — the 49.37 headline score belongs to the post-trained chat model, not the base. At the time of that review, the base checkpoint had drawn roughly fifteen downloads and eight likes. It ships as a single BF16 safetensors file with no GGUF, MLX, or other quantized builds, and it relies on custom modeling code requiring trust_remote_code, which the reviewer flagged as a maintenance liability rather than a convenience. The model card itself recommends against using the base for chat or agentic use at all, pointing those use cases to the tuned model instead — a detail easy to miss next to a press release built around agentic benchmark wins.
Why it matters beyond one release
LFM3's actual technical novelty — the d1-2B and d1-Omni-2B "decision" models, which return calibrated probability answers across text, image, and audio inputs instead of generating free text — is a genuinely different product category aimed at classification and triage pipelines rather than chatbots, and it's arriving as Google, Qwen, and Microsoft all push their own sub-3B models for on-device deployment. But the more durable story is procedural: a company can now ship something that reads as "open," distributes freely through every major model hub and runtime, and still reserve the right to charge any user who turns that openness into real revenue. For developers deciding which small model to build on, the license terms and the gap between company-reported and independently verified benchmarks matter at least as much as the parameter count on the label.