Liquid AI

Liquid AI develops Liquid Foundation Models (LFMs) and software for running AI locally on phones, laptops, and other devices. Its work combines model architecture search, hardware optimization, and training for tool use. The company progressed from specific industry applications in early 2024 to public language models, downloadable weights, mobile deployment tools, and multimodal models.[3][1][8][10][11][12]

Liquid AI's model development grew out of earlier liquid-neural-network research. That history supplies context for its input-dependent computational units and later architecture-search work.[5][10]

Ramin Hasani traced the original liquid-neural-network idea to Radu Grosu's lab at the Vienna University of Technology, where Hasani completed his PhD and Mathias Lechner his master's degree. He told TechCrunch that the work was subsequently refined and scaled in Daniela Rus's MIT CSAIL lab, with Alexander Amini and Rus joining them. Liquid time-constant networks model continuous-time dynamics through input- and state-dependent gates; later LFMs draw on this research while using different combinations of computational units.[5][10]

Liquid AI emerged from stealth on December 6, 2023, founded by Hasani, Lechner, Amini, and Rus. At launch, Hasani was CEO, Lechner CTO, and Amini chief scientific officer. TechCrunch reported a 12-person team with a presence in Boston and Palo Alto, and plans to provide private and on-premises AI infrastructure alongside a platform for customers to build their own models.[4][5]

The company's launch announcement states that it raised $46.6 million in seed capital, led by OSS Capital and PagsGroup, Stephen Pagliuca's family office. TechCrunch's contemporaneous report gives $37.5 million for a two-stage seed round and a $303 million post-money valuation. The figures differ; the company announcement was subsequently modified in August 2026.[4][5]

Joscha Bach appeared in Liquid AI's December 2023 team announcement. In the April 27, 2024 Latent Space interview, he described its focus as “very specific industry use cases” with specific customers, rather than releasing models to the general public. He identified inference, memory consumption, and training efficiency as goals, while explicitly saying, “I'm not a developer of the liquid technology.”[2][1]

On September 30, 2024, Liquid released its first series of language LFMs: dense models with 1.3 billion and 3.1 billion parameters, and a 40.3-billion-parameter mixture-of-experts model using 12 billion activated parameters. Users could try them through Liquid Playground, but the company withheld open-source release of the models to maintain its competitive position. It also identified weaknesses in zero-shot coding, precise numerical calculations, and time-sensitive information.[8]

The first LFMs used structured, adaptive operators grounded in dynamical systems, signal processing, and numerical linear algebra. Their design separates token mixing, which combines information across a sequence, from channel mixing, which combines feature dimensions. Input-dependent featurization modulates the computation. Liquid described this framework as a way to customize models for different data modalities, hardware platforms, and memory budgets.[8]

In December 2024, Liquid introduced STAR—Synthesis of Tailored Architectures. The research by Armin W. Thomas, Rom Parnichkun, Alexander Amini, Stefano Massaroli, and Michael Poli encodes architectures as hierarchical numerical “genomes.” STAR compiles and evaluates candidates, then selects and recombines successful designs. Its objectives can combine fixed properties such as parameter count and cache size with measured properties such as trained perplexity and latency on target hardware.[9]

On December 13, 2024, Liquid announced $250 million raised in Series A financing, led by AMD. TechCrunch, citing Bloomberg, reported a valuation above $2 billion. Liquid said the funding would expand compute infrastructure and accelerate edge and on-premises inference and fine-tuning stacks. AMD's Mathew Hein described a collaboration to train and deploy Liquid's models on AMD Instinct GPUs.[6][7]

Liquid released LFM2 on July 10, 2025, with downloadable dense checkpoints containing 350 million, 700 million, and 1.2 billion parameters. The stated launch license permitted academic and research use and commercial use by companies with under $10 million in revenue; companies above that threshold were directed to obtain a commercial license.[10]

For LFM2, Liquid modified STAR to evaluate more than 50 internal capability tests and directly measure peak memory, prompt-processing speed, and token-generation speed on Qualcomm Snapdragon embedded CPUs. The resulting architecture combined ten double-gated short-range convolution blocks with six grouped-query-attention blocks. Liquid tied its use of short convolutions to the target CPUs and kernel libraries optimized for those operations, making hardware performance an explicit architecture-selection criterion.[10]

Five days later, Liquid introduced LEAP and Liquid Apollo. LEAP, the Liquid Edge AI Platform, provided Android and iOS SDKs and a model library for embedding local small language models in mobile apps. Apollo was a local AI chat application that Liquid said it had acquired earlier in 2025. The announcement presented it as an iOS app supporting LEAP's model library, including LFM2, so users and developers could test models directly on their phones.[11]

On January 5, 2026, Liquid launched LFM2.5, retaining LFM2's device-optimized architecture while extending pretraining from 10 trillion to 28 trillion tokens and expanding reinforcement-learning post-training. The open-weight release included 1.2-billion-parameter Base, Instruct, and Japanese text variants, a 1.6-billion-parameter vision-language model, and a 1.5-billion-parameter audio-language model. AMD and Nexa AI were NPU launch partners.[12]

LFM2.5-Audio accepts speech and text as inputs and outputs rather than chaining separate transcription, language-model, and speech-generation stages. Its LFM-based detokenizer converts discrete tokens into audio waveforms. Liquid reported that this component was eight times faster than LFM2's Mimi detokenizer at matched precision on a mobile CPU.[12]

Liquid's March 2026 LocalCowork desktop-agent experiment tested LFM2-24B-A2B against 67 tools across 13 MCP servers. On an Apple M4 Max laptop with 36 GB unified memory and 32 GPU cores, using llama-server, Q4_K_M quantization, and flash attention, the company reported approximately 385 milliseconds per tool-selection response and 14.5 GB of memory. With temperature and top_p both set to 0.1 and a 512-token output limit, it achieved 80% single-step selection accuracy across 100 prompts, but completed only 26% of 50 three-to-six-step chains. Its proposed interaction loop let users confirm or correct a tool call before execution.[13]

On May 28, Liquid released LFM2.5-8B-A1B, a mixture-of-experts model with 8 billion total parameters and 1 billion active parameters. Compared with its October 2025 predecessor, it expanded the stated context window to 128,000 tokens and pretraining from 12 trillion to 38 trillion tokens. Liquid made the post-trained model reasoning-only, arguing that its small active parameter count made reasoning tokens inexpensive. LocalCowork switched to this model while retaining the 67-tool, 13-server menu.[17]

The June 25 LFM2.5-230M release extended the family to smaller specialized workloads. In a deliberately simple robotics demonstration, Liquid fine-tuned the model to translate instructions into skill sequences on a Unitree G1's onboard NVIDIA Jetson Orin. Commands could combine timed walking, kneeling, and pauses. The model selected pre-trained low-level skills supplied by NVIDIA's SONIC framework. Liquid recommended the checkpoint for tool use and data extraction, while advising against advanced math, code generation, and creative writing.[18]

In July 2026, Liquid published Antidoom, adapting Antislop's single-token preference-training approach to runaway repetition. It generates completions, finds repeated spans, and trains against the token beginning the first repeat, using plausible alternative next tokens as preferred choices. Final Token Preference Optimization constrains unrelated logits while allowing targeted changes; the implementation trains a LoRA adapter without requiring full correct answers. The report was written by Sam Paech, with contributions from Maxime Labonne, Justin Li, Leonie Monigatti, Nathan Ranchin, Tim Seyde, and Sergei Tilga.[14][15]

Liquid reported that Antidoom reduced repetitive looping from 10.2% to 1.4% on an early LFM2.5-2.6B checkpoint tested with hard math and coding prompts. A separate Qwen3.5-4B experiment reduced looping from 22.9% to 1% under greedy sampling. The method also required restraint: longer training could degrade performance and introduce new loops.[14]

The August 4 LFM2.5-2.6B release described a four-stage agent-training pipeline: supervised fine-tuning, domain-specialist teacher training, multi-domain on-policy distillation, and agentic reinforcement learning. During distillation, the student generates its own responses and receives token-level feedback from the teacher matched to each prompt's domain. The final stage runs multi-turn tasks inside external harnesses including Hermes Agent and OpenClaw, exposing the model to their tools and interaction patterns. Liquid retained the qualification that larger models could be preferable for complex agent tasks or coding-heavy workloads.[19]

On August 12, Liquid released LFM2.5-VL-3B, combining the 2.6B text model's pretrained base with a SigLIP2 400M NaFlex image encoder. It answers directly without emitting a reasoning trace. The release expanded screen understanding, object grounding, function calling, and multi-image input; demonstrations included labeling document-page regions, locating objects with bounding boxes, and visually navigating a website toward a requested page.[20]