othermultimodalRuns locally1.73B (LFM3-1.7B text)
LFM3: Specs, License & How to Run
Liquid AI ~2B edge family (text/VL/audio/omni/d1). Not LFM2.5-24B-A2B.
- License
- other
- Parameters
- 1.73B (LFM3-1.7B text)
- HF likes
- 160 (2026-10-28)
- First seen
- 2026-10-23
TL;DR. LFM3 is a multimodal model from Liquid AI. Parameters: 1.73B (LFM3-1.7B text).License: other. It is listed under Closed LLM APIs, ChatGPT.
| Developer | Liquid AI |
|---|---|
| Type | multimodal |
| Architecture | LFM3 tri-hybrid family branched from LFM3-1.7B-Base. Text card: 24 layers (13 gated short convolutions, 5 sliding-window attention with a 256-token window, 6 global attention); vocab 128,000. VL-2B adds an in-house vision encoder (14 layers). Audio-2B / Omni-2B add a causal audio encoder and LiquidCodec. Distinct from LFM2.5-24B-A2B (40-layer 24.0B/2.45B MoE). |
| Parameters | 1.73B (LFM3-1.7B text) |
| Modalities | text, image, video, audio, speech |
| Languages | English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish |
| Context | 256K (LFM3-1.7B); 128K (VL-2B / d1-2B); 32K (Audio-2B / Omni-2B) |
| License | other — LICENSE on LiquidAI/LFM3-1.7B is LFM Open License v1.0 (cardData.license other / license_name lfm1.0). Section 5 conditions Commercial Use on You/Legal Entity not exceeding annual revenue Threshold of USD 10,000,000. Above the threshold, Commercial Use is not licensed under this agreement. Qualified Non-Profits using the work for non-commercial or research purposes are exempt from the threshold. Same grant on the VL / Audio / Omni / d1 cards we opened. Not legal advice. |
| Hardware | Not yet verified — The LFM3-1.7B README names BF16, FP8, NVFP4, GGUF (llama.cpp / Ollama / LM Studio), MLX-4bit, and ONNX siblings. The same README says 4-bit precision “takes about 1 GB of memory.” That memory figure is vendor. forkalt has not measured peak VRAM or tok/s. (not independently tested) |
| Hugging Face likes | 160 (as of 2026-10-28) |
| Model card | https://huggingface.co/LiquidAI/LFM3-1.7B |
| First seen | 2026-10-23 |
LFM3 is Liquid AI’s ~2B-class edge family announced 27 October 2026. It is not LFM2.5-24B-A2B (the 24.0B / 2.45B text MoE already on this site). Official hosts we opened on 2026-10-28: the lead card LiquidAI/LFM3-1.7B (created 2026-10-23T21:35:42Z; last modified 2026-10-27T05:58:29Z; 160 likes / 6,311 downloads), siblings LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B, d1-Omni-2B, and LFM3-1.7B-Base, the LFM3 collection, the LFM Open License v1.0, and the launch post Introducing LFM3. API gated was false on every card in that set.
This identification page hangs off closed LLM APIs and ChatGPT alternatives. That mapping is editorial for local / on-device checkpoints. We have no sourced quality comparison against Claude, GPT, or LFM2.5. Head-to-head facts versus the Liquid page already here: LFM3 vs LFM2.5-24B-A2B. Dedicated text identification (opened 2026-11-01): LFM3-1.7B (635 likes / 32,933 downloads). Dedicated omni identification (opened 2026-10-31): LFM3-Omni-2B (227 likes / 3,080 downloads; not the text card and not LFM2.5-24B-A2B). Sibling compare: LFM3-1.7B vs LFM3-Omni-2B.
The blog says the family is “multimodal by nature” and covers text, vision, audio, omni, and decision models on one shared backbone, pre-trained on 22T tokens. Those adjectives and the 22T figure stay vendor. The text card says LFM3-1.7B is not recommended for agentic coding or complex multi-step reasoning because of its small size.
Do not confuse LFM3 d1-2B / d1-Omni-2B (this family; LFM Open License v1.0) with watch-list LiquidAI/d1-3B (LFM2.5-VL-3B fine-tune; still no dedicated page).
Family snapshot (2026-10-28)
| Card | Params (card / tensors) | Context | Pipeline | Likes / downloads | Created |
|---|---|---|---|---|---|
| LFM3-1.7B | 1.73B / 1,728,801,536 | 256K | text-generation | 160 / 6,311 | 2026-10-23 |
| LFM3-VL-2B | 1.92B (1.73B LM + 164M vision + 25M projector) / 1,918,181,888 | 128K | image-text-to-text | 67 / 255 | 2026-10-19 |
| LFM3-Audio-2B | 2B / 2,002,332,928 | 32K | any-to-any | 60 / 156 | 2026-09-08 |
| LFM3-Omni-2B | 2.17B / 2,166,928,384 | 32K | any-to-any | 68 / 83 | 2026-10-21 |
| d1-2B | 1.92B / 1,918,181,888 | 128K | image-text-to-text (decision) | 51 / 44 | 2026-10-24 |
| d1-Omni-2B | same tensors as Omni (2,166,928,384) | — | audio-text-to-text (decision) | 41 / 36 | 2026-10-24 |
| LFM3-1.7B-Base | 1.73B / 1,728,801,536 | — | text-generation | 42 / 15 | 2026-10-26 |
License on every card above: other / lfm1.0. Gated: false. Languages on the text/VL cards: EN, AR, ZH, FR, DE, IT, JA, KO, PT, ES, VI, TH, ID, HI, RU, PL. Audio/Omni speech locales on those cards add DA, EL, FI, HE, NL, SV, TR (and named English locales).
What we verified on LFM3-1.7B (2026-10-28)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/LiquidAI/LFM3-1.7B (not gated) | HF API |
| Created / last modified | 2026-10-23T21:35:42Z / 2026-10-27T05:58:29Z | same |
| Likes / downloads | 160 / 6,311 | same |
| License tag + file | other / lfm1.0 + LFM Open License v1.0 |
HF cardData + LICENSE |
| Pipeline tag | text-generation | HF API |
| Params (card) | 1.73B | README “Model details” |
| safetensors.parameters | 1,728,801,536 | HF API |
| usedStorage (BF16) | 3,475,757,943 bytes (~3.24 GiB) | same |
| Layers | 24 (13 conv + 5 SWA + 6 global) | README + blog |
| Vocab / context | 128,000 / 256K | README |
| Blog date | 27 October 2026 | liquid.ai/blog/lfm3 |
How to run (from the official README and blog)
Copied names only. forkalt has not loaded a checkpoint.
The LFM3-1.7B README lists BF16 (vLLM, SGLang, Transformers), FP8, NVFP4, GGUF (llama.cpp, Ollama, LM Studio), MLX-4bit, and ONNX. Sampling copied from that README: general/chat temperature=0.1, top_k=50, repetition_penalty=1.1; math/STEM temperature=1.0, top_p=0.95, top_k=20. The card says 4-bit “takes about 1 GB of memory.” Vendor, not our measurement.
Hosted Liquid Playground and the Liquid Console API are different artifacts from the weight grant.
Vendor benches (self-reported)
The LFM3 blog reports LiveClaw 45.60 and AutomationBench (public) 41.29 for LFM3-1.7B, versus other open models in the same table. Those rows are self-reported. They are not a Claude or GPT comparison and not a forkalt bake-off against LFM2.5-24B-A2B.
License
LICENSE is LFM Open License v1.0. Licensor is named Liquid AI, Inc. Section 5: Commercial Use rights are conditioned on You or Your Legal Entity not exceeding the Threshold of USD 10,000,000 annual revenue. Commercial Use by an entity over that threshold is not licensed under this agreement. Qualified Non-Profit Organizations using the work for Non-Commercial or Research Purposes are exempt from the threshold. Not legal advice.
Related on this site: LFM3-1.7B (dedicated text card) — LFM3-Omni-2B (dedicated omni card) — LFM3-1.7B vs LFM3-Omni-2B — LFM2.5-24B-A2B — LFM3 vs LFM2.5-24B-A2B — Spark-X2.6-4B (vs LFM2.5) — MiniCPM-V 4.7. Decision watch (different checkpoint): LiquidAI/d1-3B on classification APIs. Serving: vLLM. Catalog: home. Editorial: About.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| LiveClaw (vendor LFM3 blog) | 45.60 (self-reported) | https://www.liquid.ai/blog/lfm3 |
| AutomationBench public (vendor LFM3 blog) | 41.29 (self-reported) | https://www.liquid.ai/blog/lfm3 |
Deploy options
Comparisons
FAQ
Is LFM3 the same as LFM2.5-24B-A2B?
No. LFM3 is Liquid AI’s next generation, announced 27 October 2026: a ~2B-class family (text, vision, audio, omni, decision) on a new tri-hybrid backbone. LFM2.5-24B-A2B is the earlier 24.0B / 2.45B text MoE already on this site (card created 2026-10-12). Same LFM Open License v1.0; different generation and size. See LFM3 vs LFM2.5-24B-A2B.
Are the weights public?
Yes. Lead card LiquidAI/LFM3-1.7B is ungated (created 2026-10-23T21:35:42Z; 160 likes / 6,311 downloads in the 2026-10-28 API check). Siblings LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B, d1-Omni-2B, and LFM3-1.7B-Base were also ungated that day. Collection: huggingface.co/collections/LiquidAI/lfm3.
Can I use it commercially?
The LICENSE file is LFM Open License v1.0. Section 5 conditions Commercial Use on You/Legal Entity not exceeding USD 10,000,000 annual revenue. Above that threshold, Commercial Use is not licensed under this agreement. Qualified Non-Profits for non-commercial/research purposes are exempt from the threshold. Not legal advice.
Does it beat Claude, GPT, or LFM2.5-24B-A2B?
We do not claim that. Blog tables (LiveClaw 45.60, AutomationBench 41.29 versus other open models) are vendor self-reported. There is no sourced bake-off against named closed APIs on this page.
Sources & verification
- https://huggingface.co/LiquidAI/LFM3-1.7B
- https://huggingface.co/LiquidAI/LFM3-1.7B/raw/main/LICENSE
- https://huggingface.co/LiquidAI/LFM3-1.7B-Base
- https://huggingface.co/LiquidAI/LFM3-VL-2B
- https://huggingface.co/LiquidAI/LFM3-Audio-2B
- https://huggingface.co/LiquidAI/LFM3-Omni-2B
- https://huggingface.co/LiquidAI/d1-2B
- https://huggingface.co/LiquidAI/d1-Omni-2B
- https://huggingface.co/collections/LiquidAI/lfm3
- https://www.liquid.ai/blog/lfm3
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.