Liquid AI
Device-native foundation models.
315 followers
Device-native foundation models.
315 followers
Advanced intelligence for processors outside of data centers.
Built for the latency, privacy, and hardware constraints of the physical world.
This is the 6th launch from Liquid AI. View more
LFM3 by Liquid AI
Launching today
LFM3 is Liquid AI’s third generation of Liquid Foundation Models, with open weights and a shared backbone across text, vision, audio, omni and decision models. The first release is six 2B-class models, including a 1.7B text model that fits in about 1 GB at 4-bit and has a 256K context. Liquid says they run on phones, laptops and edge devices and can be fine-tuned on one GPU. Weights are on Hugging Face.
Launch tags:Artificial Intelligence
Launch Team
CloudflareDeploy serverless functions globally in seconds
Promoted
LFM3 is Liquid AI’s third generation of Liquid Foundation Models: small open-weight models built for running on devices, with one shared backbone across text, vision, audio, omni and decision models.
Problem: Many useful AI tasks, like local agents, on-device assistants and robotics, can’t send every request to a large cloud model. Small models are cheaper and faster to run, but they usually give up quality, and different kinds of input often need different model families.
Solution: LFM3’s first release is six 2B-class models, all branched from LFM3-1.7B-Base: a 1.7B text model, a vision-language model (the first LFM to accept video input), an audio model for speech in and out in 18 languages, an omni model that handles vision, audio and language together, and two decision models. Per Liquid, the text model fits in about 1 GB at 4-bit precision, and every model can be fine-tuned on a single GPU.
What makes it different: Liquid designed the architecture around real hardware, using a hardware-in-the-loop search across 14 device and backend combinations, including phones, a Raspberry Pi 5, laptops, a Qualcomm NPU and GPUs. Per Liquid, on-device decoding is 1.8 to 2.8 times faster than the next-fastest compared model (llama.cpp, Q4_0, 32K context), and a full 128K-token KV cache needs 1.6 GB. It was pre-trained on 22T tokens, and the text model has a 256K context.
Key features:
LFM3-1.7B for local agents, post-trained for Hermes Agent, Pi and OpenClaw
LFM3-VL-2B, which averages 71.4 across 62 vision benchmarks, within 2 points of GPT-5.6 Luna (73.4), per Liquid
LFM3-Audio-2B with time to first token as low as 50 ms
LFM3-Omni-2B for vision, audio and language in one model
Decision models (d1-2B and d1-Omni-2B) for text, images and audio
Runs on llama.cpp, MLX, vLLM and Ollama, among others
Open weights on Hugging Face, plus the Liquid AI API and Playground
Who it’s for: Developers building on-device agents, edge and robotics applications, and specialized apps where a small, fine-tunable model is enough.
P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends
1 GB at 4-bit for the text model is small enough for a phone, which is the part i care about. which phones have you actually run it on so far?
Good to see a 1.7B model post-trained for Hermes Agent. I begrudge wasting my subscription tokens for the most basic of tasks. I'll give it a go.
this looks like an absolute banger, looks like a perfect foundation to finetune a whole bunch of specific use case models
A 1.7B text model in about a gig running on a phone with 256K context is the spec that matters here, since the interesting work is going to happen on devices with no reliable connection.
The 1.7B model fitting in about 1 GB at 4-bit with a 256K context caught my eye, especially with a full 128K KV cache needing 1.6 GB. In practice on a phone, do you expect people to run the full 256K context, or is that more for laptops and edge devices where the KV cache fits more comfortably?
The hardware-in-the-loop search is thoughtful engineering - testing across Pi, Qualcomm NPU, laptops, phones. That's how you find real tradeoffs. On-device inference at 1.8-2.8x speed-up matters for latency-critical tasks. The KV cache story (1.6GB for 128K) is the real constraint - that's tight enough for voice agents and real-time systems. How's the fine-tuning community adoption looking so far?