Liquid AI LFM3
by Liquid AI
Open-weight 2B text, vision, audio and decision models that run on phones, laptops, cars and private servers
LFM3 is Liquid AI's October 2026 family of six open-weight models of about 2B parameters for text, vision, speech, omni input and structured decisions. It is for enterprises that need agents and multimodal features to run on their own devices or servers, with low latency and no data leaving the environment.
LFM3 is the third generation of Liquid Foundation Models from Liquid AI, an MIT spin-off based in Boston. Released on October 27, 2026, it is a family of six models that share one 1.7B-parameter foundation pretrained on 22 trillion tokens and then tuned for tool use and business workflows: LFM3-1.7B (text and tool calling, 256K documented context), LFM3-VL-2B (image and video input), LFM3-Audio-2B (speech in and out in 18 languages), LFM3-Omni-2B (combined vision, audio and text), and two decision models, d1-2B and d1-Omni-2B, that return structured choices with probabilities. Weights are on Hugging Face under the LFM Open License v1.0, which is free for companies under $10 million in annual revenue; larger businesses need a paid enterprise licence priced by deployment scale. The same models are sold through Liquid's hosted API, OpenRouter and Vercel AI Gateway at the same per-token rates, starting at $0.08 per million input tokens for the text model, with a 128K hosted context and zero data retention. Liquid reports that the 1.7B text model beats Qwen3.5-9B on WebArena-Verified, AutomationBench and BFCL V4 while trailing it on SWE-bench Verified, GPQA Diamond and long-context RULER; none of these figures has been independently replicated. The models run on llama.cpp, MLX, ONNX, Ollama, vLLM and SGLang, and fine-tuning fits on a single GPU. Mercedes-Benz and Shopify were pre-release testing partners, and Liquid's earlier $250 million Series A in December 2024 was led by AMD.
A platform or product engineering lead who has to ship agent, vision or voice features on devices or in a private environment where a frontier API is too slow, too costly or not allowed
Multimodal models of about 2B parameters that run locally at roughly 1-3GB of memory and can be fine-tuned on one GPU, with hosted text pricing from $0.08 per million input tokens
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based, Freemium, Contact for pricing
- Target Market
- CTOs, Enterprise Developers, ML Engineers, Product Leaders
- Deployment
- Edge-first, Open-source, API-based, Self-hosted
- Founded
- 2023
- Headquarters
- Boston, United States
Key Features
- ✓Shared 1.7B foundation across six models
Text, vision, audio, omni and decision variants share one backbone, so a team can standardise tooling and fine-tuning across modalities.
- ✓On-device inference
Runs through llama.cpp, MLX, ONNX and Ollama on laptops, phones and edge hardware; Liquid reports 43.6 tokens per second on an iPhone 17 Pro at 4-bit.
- ✓Tool calling for agents
The text model is post-trained for tool use and is designed to plug into agent frameworks such as OpenClaw, Hermes Agent and Pi.
- ✓d1 decision models
Return structured decisions with probabilities from text, images or audio, useful for routing, triage and moderation pipelines that need calibrated outputs.
- ✓Single-GPU customisation
Fine-tuning works with Liquid Finetune, TRL or Unsloth on one GPU, and companies may keep their adapted weights private.
- ✓Hosted API with zero retention
The same models are available through Liquid Console, OpenRouter and Vercel AI Gateway at identical per-token prices with zero data retention on the hosted side.
Capabilities
Use Cases
- •In-car voice assistant
An automaker runs LFM3-Audio or Omni on the vehicle so spoken commands work offline; Mercedes-Benz is exploring it for its MBUX Virtual Assistant.
- •Browser automation agent
A team uses the 1.7B text model to drive web tasks and form filling, where Liquid reports 38.42 on WebArena-Verified against 28.94 for Qwen3.5-9B.
- •Document and chart extraction at the edge
LFM3-VL-2B reads OCR and chart content on local hardware, scoring 83.9 on Liquid's OCR and chart average, keeping scans inside the network.
- •Ticket and content triage
A support or trust-and-safety team routes incoming items with a d1 decision model priced at $0.02 per million input tokens on the hosted API.
Ideal For
Best For
- ✓On-device assistants in cars, phones and laptops that must answer without a network round trip
- ✓Browser and workflow automation agents that call tools many times per task, where per-call cost matters
- ✓Private or air-gapped deployments where customer data cannot leave company infrastructure
- ✓High-volume classification and routing with the d1 decision models, which return structured choices with probabilities
- ✓Voice front ends that need fast first-token response on modest hardware
Not Ideal For
- ✗Agentic coding or complex multi-step reasoning, which Liquid's own documentation says the 1.7B model is not recommended for
- ✗Knowledge-heavy or long-context work, where the text model trails Qwen3.5-9B on GPQA Diamond and RULER 128K by Liquid's own numbers
- ✗Companies over $10 million in revenue that want to self-host without negotiating a commercial licence, since production use needs a paid enterprise agreement
Integrations
Deployment
Market Analysis
Pros
- ✓Very low hosted prices and a free self-host licence for smaller companies
- ✓Strong reported results on browser automation and tool-calling benchmarks for a 1.7B model
- ✓Broad runtime support, so the same weights run on phones, laptops and GPU servers
- ✓HN users who used earlier LFM releases for annotation called LFM3 a big step up
Cons
- ✗Liquid's own numbers show the text model trailing Qwen3.5-9B on coding, knowledge and long-context tests
- ✗HN testers saw the model insist on using tools when none were configured, and some could not load the MLX builds at launch
- ✗Hosted context is 128K against 256K for the downloadable weights, and the commercial licence threshold of $10M revenue adds a procurement step for most enterprises
- ✗All benchmark results are vendor-reported and had not been independently replicated at launch
Pricing
Open licence (under $10M revenue)
$0
- ✓Download, run and fine-tune every open LFM
- ✓Deploy on mobile, edge or on-prem
- ✓Keep fine-tunes private
- ✓Free for research, education and non-profits at any size
Hosted API
From $0.08 per million input tokens
- ✓LFM3-1.7B at $0.08 input / $0.20 output per million tokens
- ✓LFM3-VL-2B at $0.10 / $0.50
- ✓Audio and Omni at $2.00 audio input / $10.00 audio output
- ✓$2 free credit per new organisation
Enterprise licence and custom deployment
Contact for pricing
- ✓Production rights for companies over $10M revenue
- ✓OEM and on-prem deployment support
- ✓Dedicated support and SLAs
- ✓Custom architecture and optimisation
Hosted use is metered per million tokens and listed publicly, with audio billed separately at $2.00 in and $10.00 out. Self-hosting is free under the LFM Open License only for organisations under $10 million in annual revenue; above that, production use needs an enterprise licence whose price depends on deployment scale and is not published.
Security & Compliance
Sources
This page was written from 6 sources, 4 on domains other than liquid.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Recursive Flint
Apache 2.0 open-weight coding model (40B MoE, 2.5B active) built to run coding agents locally on a laptop
Arena AI Evaluations
Paid model evaluations built on the Arena leaderboard's crowdsourced human votes, for enterprises choosing an LLM
Kimi K3
2.8T-parameter open-weight MoE model with a 1M-token context, priced at $3/$15 per million tokens
TypeSafe Jev
A 'System One' decision model that returns typed choices, scores and calibrated probabilities instead of text, for agent routing and classification