Liquid AI's LFM3 Splits One 1.7B Backbone Into Six Small Models for Text, Vision, Voice and Decisions

Liquid AI’s pitch for its third model generation is that small models get useful when you specialize them, and that specialization is easier if every variant starts from the same base. On October 27, 2026, the company announced LFM3. The first release has six models of about 2 billion parameters each, covering text, vision, audio, combined “omni” input and two “decision” models. All of them branch from one base model, LFM3-1.7B-Base, and Liquid says each can be fine-tuned on a single GPU.

The base and post-trained weights are on Hugging Face, and the models are also served through Liquid’s Playground and API. The model cards list Liquid’s own lfm1.0 licence rather than a standard open-source licence such as Apache 2.0, so teams planning commercial use should read its terms. The weights went up across the week before the announcement: the vision model’s repository was created on October 19 and the base model’s on October 26.

Three kinds of layer instead of two

LFM2 combined short convolutions with attention. For LFM3, Liquid reran its hardware-in-the-loop architecture search across 14 device and backend combinations, up from one phone and one laptop CPU for LFM2. The list covers Samsung Galaxy S26 Ultra and OnePlus 15 phones, 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 Liquid calls a “tri-hybrid” layer mix. Of LFM3’s 24 layers, 13 are gated short convolutions with a kernel size of 7 (up from 3), five use sliding-window attention over 256 tokens and six use global attention. Other changes include head-wise gated attention, rotary position encoding applied to only half of each head in the global layers, and normalization on both the input and output of each block. Pre-training covered 22 trillion tokens. The last 3.5 trillion were a mid-training stage weighted toward agentic traces and tool calls, and that stage also stretched the context window to 256K tokens.

Liquid’s headline efficiency claim is that LFM3-1.7B outperforms its previous LFM2.5-2.6B with 35% fewer parameters. It also says that, per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.

Six models from one base

• LFM3-1.7B is the text model, post-trained to run inside agent harnesses such as Hermes Agent, Pi and OpenClaw.
• LFM3-VL-2B is the first LFM vision model to accept video, and it can reason step by step before answering. It comes with a new in-house vision encoder that Liquid says uses three times less compute than the SigLIP2 encoder in earlier versions.
• LFM3-Audio-2B takes text or audio and replies in text, speech or both, across 18 languages. It adds tool calling and a new codec, LiquidCodec, which Liquid says decodes speech at about 87 times real time on a single CPU thread of an Apple M4 Max.
• LFM3-Omni-2B, Liquid’s first omni model, combines vision, audio and language so users can, for example, ask spoken questions about a video and its soundtrack.
• d1-2B and d1-Omni-2B answer a list of questions about an input in one forward pass instead of generating text, each answer with a calibrated probability. New answer types let them draw boxes, extract text spans, estimate numbers and pick multiple options. d1-2B replaces the earlier d1-3B.

Liquid’s numbers

All benchmark figures come from Liquid and have not been reproduced independently. On the company’s agent and tool-use table, LFM3-1.7B averages 49.37 across the full set. Qwen3.5-9B averages 47.24 and LFM2.5-2.6B averages 31.55. The small model posts 38.42 on WebArena-Verified and 71.39 on BFCL V4, ahead of Qwen3.5-9B on both. Liquid acknowledges that “larger models still win on agentic coding and knowledge.” Qwen3.5-9B scores 51.20 on SWE-bench Verified to LFM3’s 43.40, and 80.40 on GPQA Diamond to LFM3’s 63.37.

On footprint, Liquid says a full 128K-token context needs 1.6 GB of KV cache for LFM3-1.7B, compared with 4.3 GB for Qwen3.5-4B. Across 62 vision benchmarks, the company puts LFM3-VL-2B at a 71.4 average, close to Gemma 4 12B’s 71.7.

The release also includes two worked examples of specialization. In one, Liquid fine-tuned LFM3-VL-2B on about 3,000 public photos of wind turbine blades for a few hours on one A100, which raised defect-classification accuracy to 67%. In the other, it adapted the vision model into robot-arm orchestration, navigation and vision-language-action controllers.

Day-zero support covers llama.cpp, MLX, vLLM and SGLang. Quantized GGUF, MLX, ONNX, FP8 and NVFP4 builds are already in Liquid’s Hugging Face organization.