LiquidAI on October 27 released LFM3, a family of six open-weight models built on a single 2B-parameter backbone that spans text, vision, audio, omni, and decision modalities1,3. The models are available on Hugging Face Inc., the Playground, and the Liquid AI API.
The flagship text model, LFM3-1.7B, outperforms the company's previous LFM2.5-2.6B while using 35 percent fewer parameters. All six variants branch from a shared base, LFM3-1.7B-Base, which was pre-trained on 22 trillion tokens.
A tri-hybrid architecture
LFM3 combines gated short convolutions, sliding-window attention, and global attention in what LiquidAI calls a tri-hybrid design. Of the model's 24 layers, 13 are convolutions, five are sliding-window attention with a 256-token window, and six are global attention. The architecture shares the same 128K vocabulary used by LFM2.5-8B-A1B.
A final 3.5 trillion tokens of mid-training on curated, higher-quality data extended the context window to 256K tokens. At 4-bit precision the text model fits in roughly 1 GB, and with 8K tokens in context LFM3-1.7B is approximately 25 percent faster than LFM2-1.2B. On short prompts it matches LFM2-1.2B latency across edge devices.
Benchmark results
LiquidAI published scores across a broad set of evaluations. On agentic and coding benchmarks, LFM3-1.7B scored 43.40 on SWE-bench Verified, 73.46 on LiveCodeBench v6, and 31.74 on LiveCodeBench Pro 25Q2. Tool-use scores included 71.39 on BFCL v4 and 80.36 on ToolSandbox. The model posted 45.60 on LiveClaw and 68.64 average on Claw-Eval General (EN). On long-context evaluation it scored 65.56 on RULER 128K.
For enterprise and automation tasks, LFM3-1.7B recorded 41.29 on AutomationBench, 28.20 on EnterpriseOps-Gym, and 38.42 on WebArena-Verified. Science and research scores included 28.47 on SciCode and 32.57 on MCP-Atlas.
Edge and specialization focus
Every LFM3 model can be fine-tuned on a single GPU, a design choice aimed at letting practitioners adapt models to proprietary data and deploy them from on-device to cloud. LFM3-1.7B can also be post-trained to run in agentic harnesses such as Hermes Agent, Pi, and OpenClaw.
ANALYSIS The 43.40 SWE-bench Verified score is notable for a model at the 2B-parameter scale, though direct comparisons depend on evaluation conditions that vary across labs.
The omni model is LiquidAI's first, and the vision variant now supports video in addition to images. Base and post-trained weights are available immediately under an open-weight license.