Liquid AI launches LFM3, a 2B-parameter multimodal model family with open weights

Liquid AI has announced LFM3, the third generation of its Liquid Foundation Models: a family of roughly 2-billion-parameter models that brings text, vision, audio, omni and decision models onto a single shared backbone. According to the company's announcement, which reached the front page of Hacker News, the base and post-trained models are open-weight and available now on Hugging Face, in the Liquid AI Playground and through the Liquid AI API.

One backbone, six specialized models

The first LFM3 release consists of six models, all branched from the same LFM3-1.7B-Base checkpoint and all sitting around the 2B-parameter mark. Four cover distinct input modalities — text, vision, audio and an omni model — while the remaining two are decision models specialized from the vision and omni variants. Liquid AI's argument is that developers should not run one large general-purpose model for everything; instead, they pick the variant that fits their use case, fine-tune it and run it on-device.

A tri-hybrid architecture tuned by hardware search

LFM2 showed that mixing short convolutions with attention could deliver fast on-device inference. For LFM3, Liquid AI says it redesigned that hybrid using its automated hardware-in-the-loop architecture search, targeting agentic workloads that need more capability per parameter and fast inference over long contexts.

The search also widened in scope: where LFM2 was tuned against CPUs in one phone and one laptop, LFM3 was optimized across 14 device and backend combinations, including a Samsung Galaxy S26 Ultra, a OnePlus 15, a Raspberry Pi 5, Apple M4 Max and AMD Ryzen AI laptops, Arm and x86 cloud CPUs, GPUs and the Qualcomm Hexagon NPU.

The result is what the company calls a tri-hybrid layer mix. Of LFM3's 24 layers, 13 are gated short convolutions (kernel size raised from 3 to 7), 5 are sliding-window attention with a 256-token window, and 6 are global attention. Three further changes stand out:

• Head-wise gated attention, where a sigmoid gate scales each head's output so a head can effectively switch off — which Liquid AI says helps suppress activation outliers that hurt quantization.
• Partial RoPE in global attention layers, rotating only the high-frequency half of each head so some dimensions match on content regardless of distance; sliding-window layers keep full RoPE.
• Sandwich norm, normalizing both the input and output of each block to keep activations well-scaled.

The model keeps the 128K vocabulary introduced with LFM2.5-8B-A1B, built to tokenize non-Latin scripts more efficiently. On efficiency, Liquid AI reports that LFM3-1.7B matches LFM2-1.2B on short prompts and is about 25% faster with 8K tokens in context.

Training built for agents

LFM3-1.7B-Base was pre-trained on 22T tokens, with a data mix that adds scientific papers, repository-level code, broader knowledge sources and considerably more math than LFM2's. The final 3.5T tokens form a mid-training stage on curated, higher-quality data with a larger share of agentic traces and diverse tool calls, which also extends the context window to 256K tokens.

The text model is post-trained to operate inside agent harnesses such as Hermes Agent, Pi and OpenClaw, with example tasks including inbox triage, calendar management, web research and searching for the right tool among thousands of options.

Benchmarks: strong on agents, weaker on knowledge

In the benchmark tables Liquid AI published, LFM3-1.7B averages 49.37 across the reported suites, ahead of Qwen3.5-9B at 47.24 and well ahead of predecessor LFM2.5-2.6B at 31.55. It leads on instruction following (IFEval 96.66) and function calling (BFCL V4 71.39, ToolSandbox 80.36), and it beats Qwen3.5-9B on long browser and business-app tasks such as WebArena-Verified (38.42 versus 28.94) and AutomationBench (41.29 versus 15.10).

The gaps run the other way for knowledge and hard coding: Qwen3.5-9B scores 80.40 on GPQA Diamond versus 63.37, and 51.20 on SWE-bench Verified versus 43.40. Liquid AI acknowledges that larger models still win on agentic coding and knowledge.

Hardware partners lining up

AMD and Ambarella both appear in the announcement. AMD's Josh Hort points to the combination of multimodal capability and a compact footprint as a fit for local AI on Ryzen AI processors, while Ambarella's Malhar Palkar notes that Liquid AI's LFM2.5-VL models already run on Ambarella silicon through the Cooper Model Garden, with LFM3 enablement under way on the same platform.

Why it matters

LFM3 sits at the intersection of three trends: shrinking capable models, the shift toward agentic workloads, and AI moving from the cloud to client devices. Liquid AI's bet is that a family of small, specialized, open-weight models — fine-tuned locally and run anywhere from a Raspberry Pi to a cloud CPU — will beat monolithic generalists for many real tasks, and the tri-hybrid design suggests there is still headroom in architecture search beyond the standard transformer. The caveats are familiar: these are vendor-selected benchmarks, and a 1.7B model still trails larger rivals on knowledge and coding. But for developers building on-device agents, an open, efficient starting point with a 256K-token context window is a genuinely useful addition.