LiquidAI: LFM3-1.7B

Liquid AI announced the LFM3 family on October 27, 2026, publishing ungated public Hugging Face repositories for six open-weight 2B-class models, including the text model LiquidAI/LFM3-1.7B alongside LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B and d1-Omni-2B. The subject text model carries a 256K-token combined input-and-output limit, while sibling vision and omni cards specify 32K. LFM3-1.7B uses a 1.73B-parameter, 24-layer tri-hybrid backbone combining 13 gated short-convolution layers with sliding-window and global attention, and needs roughly 1 GB of memory at 4-bit precision. Weights ship under LFM Open License v1.0, which does not license commercial use by a legal entity with at least $10 million in annual revenue.

Liquid AI announced LFM3 on October 27, 2026 as a family of six open models in the 2B parameter range, covering text, vision, audio, an omni model that sees, hears and talks, and two decision models including d1-2B. The text model LFM3-1.7B and its base variant branch from one backbone pre-trained on 22 trillion tokens with a 256K-token context window, letting teams pick the small model that fits a job and fine-tune it. This analysis reports the text model averaging 49.37 on Liquid's 29-benchmark table versus 47.24 for the much larger Qwen3.5-9B, and decoding 43.6 tokens per second on an iPhone 17 Pro using 1.90 GB of memory. It also flags that Liquid's own model card advises against using LFM3-1.7B for agentic coding or complex multi-step reasoning, and that all figures are vendor-reported.

Liquid AI launched LFM3 on October 27, 2026 as a family of six 2B-class open-weight models spanning text, vision, audio, omni, and two decision variants, all branched from a single LFM3-1.7B-Base backbone pre-trained on 22T tokens. The tri-hybrid design mixes gated short convolutions, sliding-window attention, and global attention, and the text model carries a 256K-token context window. According to this third-party analysis, which checked the released config and reference implementation, LFM3-1.7B beats the larger LFM2.5-2.6B with 35% fewer parameters and posts strong agent benchmarks including LiveClaw 45.60 and WebArena-Verified 38.42, while fitting in 1 GB at 4-bit and decoding 1.8-2.8 times faster than the next-fastest model on Liquid's tested edge devices. All benchmark figures are vendor-reported.

Qualcomm AI Hub lists LiquidAI's LFM3-1.7B as a compact on-device language model served from its model catalog, naming the artifact LiquidAI/LFM3-1.7B-GGUF and exposing a 1.05 GB Q4_0 quantization described as quantization-aware distillation. The model handles text in and text out, targets dialogue and content generation, and runs under the llama.cpp community runtime through the GenieX CLI, with the page linking the Hugging Face repository and the Liquid AI research paper. On the Snapdragon X2 Elite configuration, the listing reports a 4,096-token context window, 2,748 tokens/s prefill speed and 42.1 tokens/s decoding speed, with accuracy of 80.2% on reference versus 79.8% measured on-device. Supported hardware spans Snapdragon X Elite and X Plus 8-Core CRD, Snapdragon 8 Elite variants, and Dragonwing Q-8750, IQ-9075, IQ-X7181 and IQ-X5121 boards, plus phones such as the Galaxy S25 and S26 and Arduino VENTUNO Q. These figures are vendor-measured for this quantized build on listed Qualcomm silicon.

Benchgen's model card page for LFM3-1.7B reports vendor-supplied evaluations for Liquid AI's 1.73B-parameter text model: 71.4% on BFCL V4, 80.4% on ToolSandbox and 96.7% on IFEval, ahead of compared models up to 9B parameters on the function-calling tests. It was pre-trained on 22T tokens, with mid-training extending context to 256K tokens. The same page notes weak spots: 63.4% on GPQA Diamond trails Qwen3.5-4B and Qwen3.5-9B, and 7.2% on HLE without tools leads Liquid to advise against knowledge-heavy use. It recommends the model for local and edge agent deployments, tool calling and structured outputs, at about 1GB memory in 4-bit quantization.

Liquid AI published a first-party article on the Hugging Face blog on October 27, 2026 titled "Build specialized multimodal models with LFM3," announcing the LFM3 generation as a set of models for building specialized multimodal systems. The post was authored by the Liquid AI organization together with Hugging Face community contributors including Maxime Labonne, Anthony Maltsev, and Leonie Monigatti. The article frames LFM3 as a family for building specialized models across modalities rather than a single general-purpose checkpoint, matching the accompanying model card and documentation that list base, text, vision, audio, omni, and decision members. The scraped excerpt of the post is largely contributor metadata, so detailed technical claims are drawn from the model card and docs pages.

Liquid AI's documentation page for LFM3-1.7B confirms the model specs: 1.73B parameters, dense LFM3 tri-hybrid architecture with 24 layers split into 13 gated short convolution, 5 sliding-window attention, and 6 global attention, a 128K vocabulary, and a 256K-token context length for input and output. It is distributed under the LFM Open License v1.0. The docs give recommended generation parameters of temperature 0.1, top-k 50, and repetition penalty 1.1, with math and science runs suggested at temperature 1.0, top-p 0.95, and top-k 20. The GGUF repository's Q4_0 build is described as a quantization-aware distillation checkpoint trained to recover accuracy lost to 4-bit quantization, and quick-start paths cover Transformers, llama.cpp, vLLM, and SGLang.

Liquid AI released LFM3-1.7B as a post-trained, text-only model for local agents and specialized applications, built on the company's new tri-hybrid architecture. The checkpoint holds 1.73B parameters across 24 layers combining 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers, with a 256K-token context window and a 128,000-token vocabulary. The model is post-trained for agentic harnesses including Hermes Agent, Pi, and OpenClaw, targeting tool calling, structured outputs, and domain-specific fine-tuning rather than agentic coding or complex multi-step reasoning. At 4-bit precision it needs about 1 GB of memory, and checkpoints are published for BF16, FP8, NVFP4, GGUF, MLX, and ONNX under the LFM Open License v1.0.