Liquid AI announced LFM3 on 27 October 2026, a family of six open-weight models of about 2B parameters each, covering text, vision, audio, omni and two decision models. The company says they are available today on Hugging Face, its Playground and the Liquid AI API.
The Hugging Face model cards for all six models list the LFM Open License v1.0. The license grants commercial use only to a company that does not exceed its “Threshold”, which it defines as “annual revenue of 10 million United States dollars ($10,000,000) or more.”
Short explainer · Liquid AI LFM3
A vertical news Short narrated from the published text of this piece and related FourWeekMBA pieces.
Business Pill · On-device AI
An evergreen Business Engineer explainer on on-device AI: running a model on the phone or laptop rather than in the cloud. It explains the general idea and does not discuss Liquid AI or LFM3.
The key insight: As we read it, Liquid AI is releasing the weights while keeping a commercial line: the LFM Open License v1.0 covers commercial use only for companies below $10 million in annual revenue.
Liquid AI’s launch post says the six models “all branched from the same LFM3-1.7B-Base backbone and all around 2B parameters.” They are LFM3-1.7B for text, LFM3-VL-2B for vision, LFM3-Audio-2B, LFM3-Omni-2B, and two decision models, d1-2B and d1-Omni-2B.
The text model has a 256K-token context window. The company describes it as “a text model for local agents and specialized applications,” post-trained to run in agent harnesses such as Hermes Agent, Pi and OpenClaw.
The decision models answer questions rather than write text. Liquid AI says that given a situation and a list of questions, “they answer all of them in a single forward pass, each with a calibrated probability.”
Benchmark scores for LFM3-1.7B and Qwen3.5-9B as published in Liquid AI’s LFM3 launch post (27 October 2026); all scores were run by Liquid AI. Rows were chosen to show both the tests where the 1.7B model leads and those where the 9.7B model leads.
What Was Posted on X
Introducing ๐ป๐ต๐ผ๐น: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models. > ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ: text, 256k context window. > ๐ป๐ต๐ผ๐น-๐ ๐ป-๐ธ๐ฑ: vision-language model with video-input support for spatial understanding > ๐ป๐ต๐ผ๐น-๐ฐ๐๐๐๐-๐ธ๐ฑ: tool calls, voice chat, reasoning in 18 languages. > ๐ป๐ต๐ผ๐น-๐พ๐๐๐-๐ธ๐ฑ: an omni model combines vision, audio, and language. > ๐๐ท-๐ธ๐ฑ and ๐๐ท-๐พ๐๐๐-๐ธ๐ฑ: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input. > fine-tune on 1 GPU. > a new tri-hybrid architecture > architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs. > ๐ป๐ต๐ผ๐น-๐ ๐ป-๐ธ๐ฑ is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters. > ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ is #1 on the Pareto frontier of device intelligence ๐๐๐๐๐๐๐.๐๐๐๐๐๐.๐๐
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
Liquid AI introduced LFM3 on X on 27 October as its new generation of Liquid Foundation Models, with a new tri-hybrid architecture and new multimodal encoders. The post lists six model instances across text, audio, vision and omni, including LFM3-1.7B with a 256k context window and the d1-2B and d1-Omni-2B decision models, and says the models fine-tune on 1 GPU. This piece uses the post for the model line-up.
Liquid AI just launched LFM3, an open-weight model family whose 1.7B text model beats its 2.6B predecessor.
The gain comes from a tri-hybrid design: of 24 layers, 13 are short convolutions, 5 use 256-token sliding-window attention and 6 attend globally.
That efficiency carries into agent benchmarks run by Liquid, where the 1.7B model scores 41.29 on AutomationBench while Qwen3.5-9B manages 15.10.
Rohan Paul wrote on X on 27 October that LFM3’s 1.7B text model beats its 2.6B predecessor, that of its 24 layers 13 are short convolutions, 5 use 256-token sliding-window attention and 6 attend globally, and that on agent benchmarks run by Liquid the 1.7B model scores 41.29 on AutomationBench while Qwen3.5-9B manages 15.10. This piece uses the post as a summary of Liquid AI’s own figures.
The Architecture Behind the Speed Claims
Liquid AI calls the design tri-hybrid. The post says: “Of its 24 layers, 13 are convolutions, 5 are SWA with a 256-token window, and 6 are global attention.” SWA is sliding-window attention.
The company says LFM3-1.7B-Base was pre-trained on 22T tokens. It says that per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices, and with 8K tokens in context it is about 25% faster.
On memory, the post says “a full 128K-token context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B.” On a single Nvidia B300, it says LFM3-1.7B “can produce 3.6 billion tokens per day at 1,024 concurrent requests.”
What Liquid AI Says It Tested
KV cache needed for a full 128K-token context, as stated in Liquid AI’s LFM3 launch post (27 October 2026): 1.6 GB for LFM3-1.7B, 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B. These are the company’s own measurements. The license line is from the LFM Open License v1.0.Animated: LFM3 by Liquid AI’s own numbers, from its 27 October 2026 launch post: six open models, 3.6 billion tokens a day on one Nvidia B300 at 1,024 concurrent requests, and 1.6 GB of KV cache for a full 128K-token context.
All benchmark scores in the launch post were run by Liquid AI. On its public AutomationBench test, the company reports 41.29 for LFM3-1.7B against 15.10 for Qwen3.5-9B, a model of 9.7B parameters.
The same table shows the larger model ahead elsewhere: 51.20 against 43.40 on SWE-bench Verified, and 80.40 against 63.37 on GPQA Diamond. The post says: “Larger models still win on agentic coding and knowledge, where size matters most.”
Liquid AI’s post quotes Ivan Petrukha, a staff AI engineer at MacPaw, on a test against Qwen3.5-2B: “The smaller model won: 64% task success versus 48%, 58% fewer tokens per task.”
The Day-0 Launch Partners
The post lists day-0 launch partners: “Ambarella, AMD, Arm, Intel via OpenVINO, NVIDIA, OpenRouter, Qualcomm via Qualcomm AI Hub and GenieX, and Vercel.” It thanks Mercedes-Benz and Shopify for partnering on pre-release testing and evaluation.
Qualcomm’s Vinesh Sukumar is quoted in the post: “The next generation of AI will be more multimodal, personalized, and private while being distributed across devices and the cloud.”
Mercedes-Benz CTO Jörg Burzer is quoted saying the carmaker is “exploring how powerful on-device AI can further enhance the MBUX Virtual Assistant, while enabling fast, private and cloud-independent experiences.”
The License Line at $10M
Section 5 of the LFM Open License v1.0 says: “Any Commercial Use of the Work or a Derivative Work by a Legal Entity that exceeds the Threshold is not licensed under this Agreement.” The license sets that threshold at $10 million in annual revenue.
The license also says the threshold does not apply to a qualified non-profit’s use for non-commercial or research purposes. The documents we read do not give a price for a commercial license above the threshold.
As we read it, the launch is built around running on the device: Liquid AI says LFM3-1.7B needs 1.6 GB of KV cache for a full 128K-token context, and lists Ambarella, AMD, Arm, Intel, Nvidia and Qualcomm among its day-0 partners.
As we read it, the benchmark picture is mixed by Liquid AI’s own table: the 1.7B model leads Qwen3.5-9B on AutomationBench, 41.29 against 15.10, and trails it on SWE-bench Verified and GPQA Diamond.
As we read it, the decision models link LFM3 to a thread of launches this month: Liquid AI says d1-2B answers a list of questions in a single forward pass, each with a calibrated probability.
Liquid AI, LFM3 launch post, 27 October 2026
“Larger models still win on agentic coding and knowledge, where size matters most, but search helps close the knowledge gap, nearly doubling its score on Humanity’s Last Exam.”
Three Implications
FOR DEVELOPERS Liquid AI says the base and post-trained models are open-weight and available today on Hugging Face, its Playground and the Liquid AI API.
FOR LARGER COMPANIES The LFM Open License v1.0 says commercial use by a company at or above $10 million in annual revenue is not licensed under that agreement.
FOR CHIPMAKERS Liquid AI lists Ambarella, AMD, Arm, Intel via OpenVINO, Nvidia and Qualcomm among its day-0 launch partners.
The framework states the pattern: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Liquid AI opens the weights and keeps the large-company commercial license: the LFM Open License v1.0 covers commercial use only below $10 million in annual revenue. The framework was written in March 2026 and does not discuss Liquid AI or LFM3.
What Is Not Established
Every benchmark and speed figure in this piece comes from Liquid AI’s own launch post or from partners it quotes; we have seen no independent test of LFM3.
The launch materials give no usage, customer or revenue figures, and no price for a commercial license above the $10 million threshold.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Liquid AI released six open-weight LFM3 models of about 2B parameters on 27 October 2026, with Ambarella, AMD, Arm, Intel, Nvidia, Qualcomm and others as day-0 partners; its license covers commercial use only for companies below $10 million in annual revenue.
A note on sourcing. Model, architecture, benchmark, speed and partner details are from Liquid AI’s LFM3 launch post (27 October 2026), read in full; license terms are from the LFM Open License v1.0 published with the models on Hugging Face. All benchmark and speed figures were run by Liquid AI. Figures labelled “by our arithmetic” are ours. This is not investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
