Liquid AI released LFM3 on October 27: six open-weight models around 2B parameters, all cut from a single backbone and built to run on the device in front of you.
Liquid AI's LFM3 Family Shares One 1.7B Backbone
Liquid AI's new family is called LFM3, short for Liquid Foundation Models, third generation. Every model in the release starts from the same base, LFM3-1.7B-Base, and lands around 2B parameters.
Six models come out of that base. Four cover input types: text, vision, audio, and omni, which handles several at once. Two more are decision models, specialized from the vision and omni versions for tasks that end in an action rather than a paragraph.
The strategy is worth spelling out. Liquid AI isn't selling one big general model. It's selling a set of small ones you pick from, fine-tune for your job, and run locally. That's a different bet from the frontier labs, which keep pushing capability into larger models behind an API.
Weights for the base and post-trained versions are open and available now on Hugging Face, along with the Liquid Playground and the Liquid AI API.
The Tri-Hybrid Architecture Behind LFM3
LFM3 uses three kinds of layers in one stack, which is where the tri-hybrid name comes from. Of its 24 layers, 13 are gated short convolutions, five are sliding-window attention with a 256-token window, and six are full global attention.
The convolution kernel grew from 3 in LFM2 to 7. Sliding-window attention keeps cost flat over long inputs by looking at a nearby slice of text instead of everything. Global layers keep the model from losing the thread entirely.
A few smaller choices matter for deployment. Head-wise gating lets each attention head switch itself off when it has nothing to add, which helps when the model is compressed for a phone. Partial rotary embeddings rotate only part of each head, and a sandwich norm keeps activations in a usable range.
Liquid AI picked the mix with an automated search that runs candidates on real hardware. For LFM2 that meant one phone and one laptop. For LFM3 it covers 14 device and backend combinations, including a Samsung Galaxy S26 Ultra, a OnePlus 15, a Raspberry Pi 5, Apple's M4 Max, AMD's Ryzen AI, Arm and x86 cloud CPUs, GPUs, and Qualcomm's Hexagon NPU.
LFM3 Training: 22T Tokens and a 256K Context
The base model saw 22 trillion tokens of pre-training data. Compared with LFM2, Liquid AI says the mix added more scientific papers, repository-level code, broader knowledge sources and considerably more math.
The last 3.5 trillion tokens were a mid-training stage on curated data with more agentic traces and tool-calling examples, which is preparation for the reinforcement learning that comes later. That stage also stretched the context window to 256K tokens.
Post-training ran in four stages: supervised fine-tuning on 1.75 trillion tokens, hybrid reinforcement learning, distillation from six specialist RL teachers into one policy, and then agentic RL. The vocabulary holds 128K entries, carried over from LFM2.5-8B-A1B, and Liquid AI says it handles non-Latin scripts more efficiently than an English-first vocabulary would.
What Liquid AI Claims About LFM3 Speed and Memory
Every number in this section comes from Liquid AI. None of it has been checked by an outside lab yet, and the company's model cards are the source.
Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices, and with 8K tokens of context it runs about 25% faster. On three devices the company tested, it was the fastest model overall, decoding 1.8 to 2.8 times faster than the runner-up while using the least memory.
The memory claim is the one worth repeating. Liquid AI says a 128K context needs 1.6 GB of KV cache, the working memory a model keeps for the text it's reading. It puts Qwen3.5-4B at 4.3 GB and MiniCPM5-2B at 5.6 GB for the same job. If that holds on your hardware, it changes which models fit on a laptop.
For serving, Liquid AI reports 3.6 billion tokens per day through 1,024 concurrent requests on a single NVIDIA B300 GPU using SGLang in BF16. That's a data-center figure rather than an edge one, and it's still self-reported.
LFM3 Benchmarks Come With a Company Asterisk
The benchmark table is where enthusiasm should slow down. Liquid AI reports BFCL V4 at 71.39 and ToolSandbox at 80.36 for instruction following and function calling. It says the family beats Qwen3.5-9B on long-horizon tasks in browsers and business apps, measured on WebArena-Verified and AutomationBench.
MarkTechPost summarized the company's own numbers as an average of 49.37 across 29 benchmarks for LFM3-1.7B, against 47.24 for Qwen3.5-9B. A 2B model matching a 9B one on an average is a strong result, and an average hides which tests went badly.
Liquid AI says as much itself. On agentic coding and knowledge-heavy tasks, larger models still win. That boundary is the useful part of the disclosure. A small model that's competitive on tool calls and instruction following is a good fit for a phone assistant. It isn't a replacement for a frontier model on a hard research question.
Who Should Fine-Tune LFM3, and Who Shouldn't
LFM3 makes sense if you have a narrow job, your data can't leave the building, and latency matters. A retailer sorting support tickets on local servers, a clinic classifying intake forms, a hardware vendor shipping an offline voice assistant: those are the profiles the family was built for.
Skip it if you need broad world knowledge or heavy reasoning. Skip it too if you can't evaluate your own fine-tune, because an open-weight model is only as good as the tests you build around it.
The family is available now, and the weights are free to download. The real test comes in a few months, when independent benchmarks and shipping products tell us whether 2B parameters on one backbone was enough.
