LFM3: what Liquid AI released, how it scores, and where to run it
Liquid AI’s LFM3 is six open-weight ~2B models, including LFM3-1.7B (256K context) and LFM3-Omni-2B (text, image, video and audio in; text or speech out). Specs, benchmarks, license and API price.
Oct 30, 2026 · API ManyLiquid AI released LFM3 on October 27, 2026. It is the third generation of its Liquid Foundation Models and the first release covers six models in the 2B-parameter range, all branched from one backbone, LFM3-1.7B-Base, pre-trained on 22 trillion tokens. Liquid positions them as small models you fine-tune for one job and run on the device: the text model fits in about 1 GB at 4-bit precision.
Sources:Liquid AI: Introducing LFM3 (Oct 27, 2026) · Liquid AI on Hugging Face: Build specialized multimodal models with LFM3
The LFM3 family at a glance
| Model | Input | Output | Parameters |
|---|---|---|---|
| LFM3-1.7B | Text | Text | 1.73B |
| LFM3-VL-2B | Text, images, video | Text | 1.9B |
| LFM3-Audio-2B | Text, audio | Text, speech | ~2B |
| LFM3-Omni-2B | Text, images, video, audio | Text, speech | 2.17B |
| d1-2B | Text, images, video | Calibrated answers (decisions) | ~2B |
| d1-Omni-2B | Text, images, video, audio | Calibrated answers (decisions) | ~2B |
Sources:Liquid AI on Hugging Face: Build specialized multimodal models with LFM3 · Hugging Face: LiquidAI/LFM3-1.7B model card · Hugging Face: LiquidAI/LFM3-VL-2B model card · Hugging Face: LiquidAI/LFM3-Omni-2B model card
LFM3-1.7B: the text model
| LFM3-1.7B | Source | |
|---|---|---|
| Released | October 27, 2026 | Liquid AI |
| Parameters | 1.73B (dense) | Model card |
| Architecture | 24 layers: 13 gated short convolutions, 5 sliding-window attention, 6 global attention | Model card |
| Context window | 256K tokens (input + output) | Liquid docs |
| Languages | 16, including English, Chinese, Japanese, Korean, Arabic, Hindi and major European languages | Model card |
| Reasoning | On by default; can be turned off with enable_thinking=False | Model card |
| Memory | About 1 GB at 4-bit precision | Liquid AI |
| License | LFM Open License v1.0 (open weights) | Liquid docs |
| API price (Liquid) | $0.08 input, $0.02 cached input, $0.20 output per 1M tokens | Liquid pricing |
| On API Many | No | See the note above |
Liquid post-trained LFM3-1.7B for agent harnesses such as Hermes Agent, Pi and OpenClaw, with native tool calling and deferred “tool search”. It also says plainly what the model is not for: agentic coding and complex multi-step reasoning, and knowledge-heavy questions without retrieval or search.
Source:Hugging Face: LiquidAI/LFM3-1.7B model card
What Liquid AI reports for LFM3-1.7B
| Benchmark (Liquid AI’s evaluation) | LFM3-1.7B | Qwen3.5-4B | Qwen3.5-9B |
|---|---|---|---|
| BFCL V4 (function calling) | 71.39 | 55.12 | 62.83 |
| IFBench (instruction following) | 73.03 | 38.16 | 48.84 |
| WebArena-Verified (browser agents) | 38.42 | 21.92 | 28.94 |
| AutomationBench (public) | 41.29 | 8.67 | 15.10 |
| SWE-bench Verified (OpenCode) | 43.40 | 41.20 | 51.20 |
| GPQA Diamond | 63.37 | 75.76 | 80.40 |
| RULER 128K | 65.56 | 84.63 | 87.11 |
| Average of all 29 scores | 49.37 | 39.27 | 47.24 |
Sources:Hugging Face: LiquidAI/LFM3-1.7B model card · Liquid AI: Introducing LFM3 (Oct 27, 2026)
LFM3-Omni-2B: one model that sees, listens and talks
- Inputs: text, images, video with or without its soundtrack, and 16 kHz audio. Video is sampled at 2 frames per second.
- Outputs: text, 24 kHz speech, or interleaved text and speech, so one model can replace a separate speech recognizer, vision-language model, LLM and text-to-speech model.
- Context: 32K tokens (input + output). It is a non-reasoning model.
- Size: 2.17B parameters, designed for phones, laptops and other edge hardware.
- License: LFM Open License v1.0. Liquid’s pricing page lists no API price for it, so in practice you run the weights yourself.
Sources:Hugging Face: LiquidAI/LFM3-Omni-2B model card · Liquid AI: Pricing
| Benchmark (Liquid AI’s evaluation) | LFM3-Omni-2B | Qwen2.5-Omni-7B | Gemma 4 E4B |
|---|---|---|---|
| Daily-Omni (video + audio) | 70.1 | 61.3 | 54.1 |
| WorldSense (video + audio) | 47.1 | 45.4 | 43.5 |
| AV-SpeakerBench | 46.8 | 43.2 | 36.4 |
| OmniBench overall (image + audio) | 46.3 | 49.7 | 20.7 |
| RealWorldQA (spoken visual QA) | 64.7 | 69.9 | 49.0 |
Vision, audio and decision models
- LFM3-VL-2B adds video input and step-by-step thinking. Across 62 vision benchmarks Liquid reports an average of 71.4, against 70.0 for Gemini 3.5 Flash Lite and 73.4 for GPT-5.6 Luna.
- LFM3-Audio-2B handles speech recognition, speech synthesis and translation in 18 languages, plus audio understanding and tool calling from spoken requests. Liquid reports 6.2% English word error rate.
- d1-2B and d1-Omni-2B are decision models: given an input and a list of questions, they answer all of them in one forward pass, each with a calibrated probability. d1-2B replaces the earlier d1-3B.
Source:Liquid AI: Introducing LFM3 (Oct 27, 2026)
Where you can run LFM3 today
- Open weights on Hugging Face in BF16, FP8, NVFP4, GGUF, MLX and ONNX, with support in llama.cpp, MLX, vLLM, SGLang, Ollama and LM Studio. Every model can be fine-tuned on a single GPU.
- Liquid API: LFM3-1.7B costs $0.08 per 1M input tokens, $0.02 cached and $0.20 output. New Liquid Console organizations get $2 of free credit.
- Liquid Playground: an online chat for trying the models without setup.
- OpenRouter and Vercel are named as day-0 launch partners, but as of October 30, 2026 neither public model list includes an LFM3 model; OpenRouter lists LFM2.5 models from Liquid.
- License: the LFM Open License is free for commercial use until a company passes $10 million in annual revenue; above that Liquid asks for a paid license.
Sources:Hugging Face: LiquidAI/LFM3-1.7B model card · Liquid AI: Introducing LFM3 (Oct 27, 2026) · Liquid AI: Pricing · OpenRouter: Liquid models · Vercel AI Gateway: models
What to use on API Many instead
If you need a small, cheap hosted model behind one OpenAI-compatible endpoint rather than weights to run yourself, API Many offers GPT-5.6 Luna from $0.20 in · $1.20 out, Gemini 3.5 Flash Lite from $0.30 in · $2.499 out and GPT-6 Luna from $0.10 in · $0.50 out. The first two are the hosted models Liquid compares LFM3-VL-2B against.
Related
- LFM3-1.7B vs LFM2.5-2.6BWhat changed from the previous Liquid text model, side by side.
- Spark-X2.6-4B: the 4B open-weight Spark model, explainedAnother small open-weight model built for on-device agents.
FAQ
What is LFM3?
LFM3 is Liquid AI’s third generation of Liquid Foundation Models, released October 27, 2026. The first release has six 2B-class models for text, vision, audio, omni (all modalities) and decisions, all built on the LFM3-1.7B-Base backbone.
What is LFM3-Omni?
LFM3-Omni-2B is Liquid’s first omni model: one 2.17B-parameter model that takes text, images, video and audio and answers in text, speech or both, with a 32K-token context.
Is LFM3 open source?
The weights are open and downloadable under the LFM Open License v1.0. It is not a standard open-source license: commercial use is free until your company passes $10 million in annual revenue.
How much does the LFM3 API cost?
Liquid’s API charges $0.08 per 1M input tokens, $0.02 per 1M cached input tokens and $0.20 per 1M output tokens for LFM3-1.7B. Liquid lists no API price for the other LFM3 models as of October 30, 2026.
What is LFM3-1.7B’s context window?
256K tokens, input and output combined. LFM3-Omni-2B has 32K.
Can I use LFM3 on API Many?
Not currently. API Many does not offer LFM3. For a small hosted model on API Many, try GPT-6 Luna, GPT-5.6 Luna or Gemini 3.5 Flash Lite.
Sources
- Liquid AI: Introducing LFM3 (Oct 27, 2026) — Release date, family overview, all benchmark tables, launch partners.
- Liquid AI on Hugging Face: Build specialized multimodal models with LFM3 — Model list with inputs and outputs.
- Hugging Face: LiquidAI/LFM3-1.7B model card — LFM3-1.7B specs, intended use, sampling and full benchmark table.
- Liquid docs: LFM3-1.7B — LFM3-1.7B context window and license.
- Hugging Face: LiquidAI/LFM3-Omni-2B model card — LFM3-Omni-2B specs, inputs, outputs and context.
- Hugging Face: LiquidAI/LFM3-VL-2B model card — LFM3-VL-2B model card.
- Liquid AI: Pricing — Liquid API prices and license terms.
- Liquid docs: LFM Open License — LFM Open License.
- OpenRouter: Liquid models — Liquid models listed on OpenRouter (no LFM3 as of Oct 30, 2026).
- Vercel AI Gateway: models — Vercel AI Gateway model list (no LFM3 as of Oct 30, 2026).
Every model in this post is available on API Many. Pay as you go from $10.