Liquid AI LFM3 Open-Weight Models Explained: Six 2B Models That Put AI Agents on Your Phone (World Tech)
Illustrative image: a smartphone in hand (Wikimedia Commons, CC BY 2.0). This is NOT a Liquid AI product image — LFM3 is a model family, not a device; the photo illustrates the on-device AI the models target.
TL;DR: Liquid AI LFM3 open weight models explained: Liquid AI's six new 2B-class models share one 1.7B backbone, run AI agents on phones and laptops, and are open-weight on Hugging Face — but their headline benchmarks are the company's own, not independent tests.
📌 KEY HIGHLIGHTS
- Six models, one backbone: text, vision, audio, omni and two decision models, all branched from the same LFM3-1.7B-Base — announced 27 October 2026.
- Phone-sized: the 1.7B text model fits in about 1 GB at 4-bit precision, with a 256K-token context window.
- Agent-first: Liquid AI reports 38.42 on WebArena-Verified, beating Qwen3.5-9B's 28.94 — but larger models still win on coding and knowledge.
- Open-weight, with a catch: free commercial use under the LFM Open License v1.0 below $10 million annual revenue; bigger companies need a separate agreement.
- Available now: weights on Hugging Face, plus the Liquid AI Playground and API, with GGUF, MLX, ONNX, FP8 and NVFP4 builds.
Liquid AI LFM3 open weight models explained: the six models
Liquid AI announced LFM3 on 27 October 2026 as the new generation of its Liquid Foundation Models. Instead of one large general-purpose model, the family gives developers six specialists of around 2 billion parameters each, all built on the same LFM3-1.7B-Base backbone. The idea is simple: pick the model that fits the job, fine-tune it on your own data on a single GPU, and run it where your application lives — on a phone, a laptop, a private server or the cloud — rather than paying a cloud API bill for every request.
The backbone itself is a new "tri-hybrid" design: 24 layers mixing 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers, found through a hardware-in-the-loop search across 14 device and backend combinations — including a Samsung Galaxy S26 Ultra, a OnePlus 15, a Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI and the Qualcomm Hexagon NPU. The base model was pre-trained on 22 trillion tokens, with a final 3.5-trillion-token stage focused on agentic traces and tool calls that also stretched the context window to 256K tokens.
| Model | What it does | Inputs → outputs |
|---|---|---|
| LFM3-1.7B | Text model for local agents (inbox triage, calendar, web research, WhatsApp briefings) | Text → text, 256K context |
| LFM3-VL-2B | Vision model — the first LFM vision model to take video input | Text, image, video → text |
| LFM3-Audio-2B | Audio model with tool calling across 18 languages | Text or audio → text or speech |
| LFM3-Omni-2B | Liquid AI's first omni model — sees, listens and talks back | Text, image, video, audio → text and speech |
| d1-2B | Decision model built on the vision model — answers in one forward pass | Text, pictures → typed answers with probabilities |
| d1-Omni-2B | Decision model built on the omni model — the same decisions by voice | Text, pictures, audio → typed answers with probabilities |
Illustrative image: server racks at NERSC (Wikimedia Commons, CC0). LFM3's pitch is the opposite direction — moving everyday agent work off servers like these and onto the device itself.
What the benchmarks actually show — and what they don't
On Liquid AI's own numbers, LFM3-1.7B is a specialist, not a giant-killer. It leads on agent and tool-use tests — WebArena-Verified 38.42 against 28.94 for Qwen3.5-9B and just 3.02 for its predecessor LFM2.5-2.6B, AutomationBench 41.29 against 15.10, and BFCL V4 function calling 71.39 against 62.83. It also decodes 1.8 to 2.8 times faster than the next model Liquid AI tested on-device, and a 128K-token context needs only 1.6 GB of key-value cache against 4.3 GB for Qwen3.5-4B.
| Benchmark (Liquid AI's own tests) | LFM3-1.7B | Qwen3.5-9B |
|---|---|---|
| WebArena-Verified (browser agents) | 38.42 | 28.94 |
| AutomationBench, public (business apps) | 41.29 | 15.10 |
| BFCL V4 (function calling) | 71.39 | 62.83 |
| GPQA Diamond (knowledge) | 63.37 | 80.40 |
| SWE-bench Verified (agentic coding) | 43.40 | 51.20 |
| RULER 128K (long context) | 65.56 | 87.11 |
The other side of the table matters just as much. Larger models still win on agentic coding and knowledge, where size matters most — Liquid AI says so itself, and its own model card does not recommend LFM3-1.7B for agentic coding or complex multi-step reasoning. And every number above is vendor-reported: Liquid AI ran the evaluations internally, and no independent third-party reproduction had been published at announcement time. Treat the 12.7x agentic jump over LFM2.5 as a claim to test on your own workload, not a settled fact. The vision model tells a similar story: LFM3-VL-2B averages 71.4 across 62 vision benchmarks, matching Google's Gemma 4 12B despite being six times smaller — again, on Liquid AI's own evaluation.
The decision models are the most unusual part of the release. Instead of writing an answer, d1-2B and d1-Omni-2B decide in milliseconds: give them a support ticket, invoice, app screen, photo or voice note plus a list of questions, and they answer all of them in a single forward pass, each with a calibrated probability. New typed answers can point to a box in an image, read text and dates out of a document (saying "not found" when a field is missing), estimate a number with a range, or pick every option that applies. The smaller d1-2B beats its predecessor d1-3B on Decision Bench (85.2 vs 83.6) with a third fewer parameters.
Why it matters for India readers
For Indian developers and startups, LFM3's real appeal is cost and control, not benchmark bragging rights. A model that fine-tunes on a single GPU and runs on a phone or laptop means an AI feature without a large cloud bill — and without sending user data to a server. That matters for privacy-sensitive work in health, finance and customer support, and for apps that must work on patchy connectivity. The audio model's 18 languages and the vision model's document-reading strength also fit India's multilingual, document-heavy workflows, though Liquid AI has announced no India-specific partnership, pricing or language list for Indian languages — so treat any India claim beyond the general one with care.
The timing fits a wider shift. Days earlier, OpenAI's GPT-6.1 Terra halved the price of computer-use AI, and Google DeepMind's Genie 4 showed world models you can steer live — the industry is racing in both directions at once: bigger frontier models in the cloud, and smaller specialists on the device. LFM3 is firmly in the second race, alongside Google's Gemma and Alibaba's Qwen small models. And the cautionary tale from this week still stands: one Qwen model was caught trying to route requests through its maker's cloud — "on-device" is only as private as the implementation, so developers should verify what actually leaves the device.
Should you care?
✅ Try it if you build local agents, voice tools or document workflows and want a small, fine-tunable, open-weight model to test this week — it is free to download from Hugging Face. ⏳ Wait if you need heavy coding, deep knowledge work or proven production numbers — run your own evaluation first, and remember the licence threshold if your revenue is above $10 million.
FAQs
What is Liquid AI LFM3? LFM3 is Liquid AI's new family of six open-weight models in the 2B parameter range, announced 27 October 2026, all branching from the same LFM3-1.7B-Base backbone: text, vision, audio, omni and two decision models. It is built for specialised, on-device intelligence rather than general-purpose chat.
Can LFM3 run on a phone? Liquid AI says the 1.73B text model takes about 1 GB of memory at 4-bit precision, and its architecture search covered phones, a Raspberry Pi 5, laptops and NPUs. That is a footprint claim, not a promise that every phone runs every model well — the full app, context and tools will need more memory, and the omni and vision models are heavier.
Is LFM3 free for commercial use? Under the LFM Open License v1.0, commercial use is free below $10 million in annual revenue; companies at or above that threshold need a separate commercial agreement with Liquid AI. It is a custom licence, not Apache 2.0, so read its terms before shipping a product on it.
Is LFM3 better than Qwen3.5-9B? On Liquid AI's own agent and tool-use tests, yes — WebArena-Verified 38.42 vs 28.94 and BFCL V4 71.39 vs 62.83. On knowledge and coding tests, no — Qwen3.5-9B leads GPQA Diamond 80.40 vs 63.37 and SWE-bench Verified 51.20 vs 43.40. No independent reproduction of Liquid AI's numbers had been published at announcement time.
What are LFM3 decision models, and what can they do? d1-2B and d1-Omni-2B answer questions instead of generating text: they return typed, calibrated answers — yes/no, choices, scores, bounding boxes, extracted text and dates, or estimated ranges — for many questions in a single forward pass, in milliseconds, on-device. d1-Omni-2B takes the same questions by voice.
Where do I get LFM3? Open weights are on Hugging Face, plus the Liquid AI Playground and API, with GGUF, MLX, ONNX, BF16, FP8 and NVFP4 checkpoint variants for llama.cpp, MLX, vLLM, SGLang and ONNX Runtime. Day-0 launch partners include AMD, Arm, NVIDIA, Qualcomm and Vercel, with Mercedes-Benz and Shopify testing the models before release.
Takeaway: LFM3 is not a frontier-model rival — it is a bet that small, specialised, on-device agents are the next useful thing. The benchmarks are promising and honestly caveated by their own maker; the real test is what developers build with them in the next few months.
Follow World Tech on TechByCountry for tomorrow's 6 PM update.
📸 Get tech in the last 24 hours on Instagram — follow @_techinlast24hrs for daily posts (9 AM–9 PM) and reels at 12:30 PM & 8:30 PM.
Sources
- Liquid AI — Introducing LFM3: Every Modality for Specialized Intelligence (primary source, 27 October 2026)
- VentureBeat — Liquid AI's new LFM3 models put browser agents, voice commands, and image recognition onto enterprise-controlled devices
- MarkTechPost — Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models
- Hugging Face — LiquidAI/LFM3-1.7B model card (benchmark table and evaluation details)
- TechTimes — Liquid AI Releases LFM3: Six Open-Weight Models Fuse Text, Vision, Audio on One 2B Backbone (28 October 2026)
Comments
Post a Comment