Liquid AI

Today, we release LFM2.5, our most capable family of tiny on-device foundation models. It’s built to power reliable on-device agentic applications: higher quality, lower latency, and broader modality support in the ~1B parameter class. > LFM2.5 builds on our LFM2 device-optimized hybrid architecture > Pretraining scaled from 10T → 28T tokens > Expanded reinforcement learning post-training > Higher ceilings for instruction following 🧵

AI operating at the edge: space! 👌🏻

AI is beginning to move beyond the clouds… Registration is open for Hack #05: AI in Space, in collaboration with @DPhiSpace. A hackathon exploring what becomes possible when AI operates closer to satellites, orbital systems, and space-based data. For developers, researchers, and builders interested in the future of AI in space. Register → luma.com/n9cw58h0 Learn more → hackathons.liquid.ai 🚀 Join the conversation → discord.com/channels/13854…

@biogerontology 🙏🏻

@liquidai rocks

a vision language model too fast for human eyes! kudos @xenovacom 🐐

Fantastic discussion between my friends Alex and Ramin on why the collaboration between @liquidai and @InsilicoMed is so important in advancing BioAI! Alex also demos real-life examples of using AI models in drug discovery! No doubt that AI will make the world a better place!

Just got this set up on my 2024 M3 MacBook Air with 24Gb of RAM. MLX build. 12Gb total memory footprint with 1-4 second response time for simple prompts. This memory footprint is insane! I didn’t expect any local model to even work half decent on this Mac. Well done @liquidai

> 385ms average tool selection. > 67 tools across 13 MCP servers. > 14.5GB memory footprint. > Zero network calls. LocalCowork is an AI agent that runs on a MacBook. Open source. 🧵

A 24-billion-parameter model just ran on a laptop and picked the right tool in under half a second. The real story is that tool-calling agents finally became fast enough to feel like software. Liquid built LFM2-24B-A2B using a hybrid architecture that mixes convolution blocks with grouped query attention in a 1:3 ratio. Only 2.3 billion parameters activate per token, even though the full model holds 24 billion. That sparse activation pattern is why it fits in 14.5 GB of memory and dispatches tools in 385 milliseconds on an M4 Max. The architecture was designed through hardware-in-the-loop search, meaning they optimized the model structure by testing it directly on the chips it would run on. No cloud translation layer. No API roundtrip. The model, the tools, and your data stay on the machine. This unlocks three things that were impractical before: 1. Regulated industries can run agents on employee laptops without data leaving the device. 2. Developers can prototype multi-tool workflows without managing API keys or rate limits. 3. Security teams get full audit trails without vendor subprocessors in the loop. The model hit 80% accuracy on single-step tool selection across 67 tools spanning 13 MCP servers. If this performance holds at scale, two assumptions need updating. First, on-device agents are no longer a battery-life trade-off; they're a compliance feature. Second, the bottleneck in agentic workflows is shifting from model capability to tool ecosystem maturity.

watch this to see what we are bringing to the next generation of PCs!

When latency, memory, and privacy are real constraints, they decide whether your agent is a product or just a demo. That’s where LFM2-24B-A2B shines. Read the blog: liquid.ai/blog/no-cloud-… LocalCowork is open source and available in our Cookbook: github.com/Liquid4All/coo…

An example workflow the agent executed locally: search receipts → OCR → parse vendor/date/items → check duplicates → export CSV → flag anomalies → generate reconciliation report. A real workflow. Running entirely on consumer hardware. youtu.be/WnxxW2jTDgE

We’re teaming up with @InSilicoMed to create lightweight scientific foundation models for pharmaceutical research. Together, we are building a series of liquid foundation models with state-of-the-art performance across multiple drug discovery subdomains. 💊 Our goal is to push the frontier of drug discovery beyond single-purpose, specialized models and towards foundational generalist models that are useful and capable of ingesting proprietary molecules, assays, and target data entirely within local private instances. The first model in line is LFM2-2.6B-MMAI, a small model that achieves cloud-scale performance while operating entirely on private infrastructure: > Molecular optimization: Up to 98.8% success on MuMO-Instruct multi-parameter optimization. > Affinity prediction: Outperformed GPT-5.1, Claude Opus 4.5, and Grok-4.1 on Insilico’s 2.5M / 689-target benchmark. > Chemical reasoning: Strong functional-group reasoning (FGBench) and solid 1-step retrosynthesis (ChemCensor). By combining Liquid AI’s efficient LFM technology with Insilico’s MMAI Gym, a comprehensive training platform w/over 1,000 pharmaceutical benchmarks, we observe that on-premise deployment can deliver competitive results across the full spectrum of drug discovery tasks, all in a single system. These capabilities unlock immediately useful applications for pharmaceutical companies, particularly in high-frequency ADMET screening, medicinal chemistry-facing lead optimization, and retrosynthesis feasibility assessment that prevents wasted experimental effort.