Liquid AI launches LFM3: six open 2B models from one shared base model

Liquid AI launches LFM3: six open 2B models from one shared base model

On 27 October 2026, Liquid AI released a new generation of Liquid Foundation Models: six open 2B-class models for text, vision, audio, omni and decisions, all branched from a single shared base model.

Liquid AI released a new generation of Liquid Foundation Models on 27 October 2026: six open 2B-class models for text, vision, audio, omni and decision tasks, all branched from a single shared base model. The text model is claimed to require around 1 GB of memory at 4-bit compression with 256,000 tokens of context — enough to run on a phone or laptop without a dedicated GPU. But all benchmark figures are the company's own, and the license carries a revenue cap of $10 million.

The news: six models, one base

LFM3 is a family of six open-weight models, all in the 2B class and all built on the same shared backbone, according to MarkTechPost. The family covers text, vision, audio, omni and decision tasks, and each model is said to be fine-tunable on a single GPU (MarkTechPost).

All six models branch from one and the same pretrained backbone, LFM3-1.7B-Base, which Liquid has also released (Orca Router). That is the whole point of the setup: rather than training each modality from scratch, the text, vision, audio, omni and decision models share the same foundation.

The models range from the text model LFM3-1.7B with 1.73 billion parameters and 256K context, via LFM3-VL-2B with 128K context, to LFM3-Omni-2B with 2.2 billion parameters (MarkTechPost). The omni model accepts text, image, video and audio as input, and responds with text or speech (Orca Router).

The tri-hybrid architecture, explained

The backbone mixes three operator types across 24 layers: 13 gated short convolutions, five sliding-window attention layers with a 256-token window, and six layers of full global attention (MarkTechPost). The convolution kernels have grown from 3 to 7 (MarkTechPost). The setup replaces the LFM2 generation's two-way mixing with three layer types (Orca Router).

According to the company, the architecture choice was not purely manual, but the result of a hardware-in-the-loop search across 14 combinations of devices and backends, including a Raspberry Pi 5 and a Qualcomm Hexagon NPU (Startup Fortune). The point of the method, per the company, is that model candidates are evaluated on real hardware during the search, so the architecture is selected based on how efficient it actually is on devices without a GPU — not just on benchmark scores.

A common misreading: 256 tokens is not the context

The architecture table has caused confusion. LFM3's sliding-window attention layers operate with a 256-token window — but that is the attention geometry per layer, not the model's context length. The text model takes 256,000 tokens of context (Orca Router). Readers who interpreted the layer table as a 256-token memory have misread the model: the window governs how far each individual attention layer "sees" locally, while the global layers and training make the full 256K context available.

It is precisely this combination of small window and long context that also makes the model memory-friendly. Liquid reports that a full KV cache at 128K tokens takes 1.6 GB, versus 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B on the same measurement (Orca Router, citing Liquid's own figures).

The numbers — all from Liquid AI itself

According to Liquid's own evaluation, the text model averages 49.37 across 29 benchmarks, versus 47.24 for Qwen3.5-9B — that is, a model roughly one-fifth the size (MarkTechPost, which explicitly flags the comparison as Liquid's own eval).

The vision model LFM3-VL-2B is said to average 71.4 across 62 benchmarks, up from 57.6 for its predecessor LFM2.5-VL-3B — despite having fewer parameters. It is, according to the coverage, Liquid's first model to accept video input (Startup Fortune, citing the company's own figures).

The audio model LFM3-Audio-2B supports 18 languages and scores 3.83 on the speech quality measure UTMOSv2, versus 3.22 for LFM2.5-Audio-1.5B. Its causal encoder is said to reduce time to first token to as low as 50 milliseconds (Startup Fortune, again Liquid's own reported figures).

An important caveat: none of these comparisons has been independently verified. No independent evaluation of LFM3 exists in the source material, and the outlets that have covered the launch likely build on the same company announcement and technical documentation.

The training: 22T tokens and an agent-heavy final stretch

The base model was trained on 22 trillion tokens. The final 3.5 trillion were an agent-heavy mid-training phase that also extended the context to 256K (MarkTechPost). Post-training ran across four stages: first SFT on 1.75T tokens, then mixed RL, then distillation of six specialist teacher models into one model, and finally agentic RL (MarkTechPost).

This training profile partly explains why the models are oriented toward agentic tasks: significant resources late in training were spent making them usable as tools that call functions and carry out multi-step tasks, not just as chatbots.

The license: not Apache 2.0

LFM3 is released under the LFM Open License v1.0, which has a revenue cap of $10 million (MarkTechPost). That distinguishes the family from competitors in the same class — MiniCPM5-2B and Qwen3.5-4B are both released under Apache 2.0, which has no such cap (MarkTechPost). For researchers and smaller companies this makes little practical difference, but commercial actors that cross the threshold must verify the terms against the full license text before building production systems on the model.

Open questions

Several things remain unclear:

• No primary source in the public record. Liquid's own announcement, technical report and model cards are not available in the sources AIMag has had access to; all specifications and figures are relayed through secondary coverage that likely builds on the same company material. The benchmark claims should therefore be read as company claims until independent evaluation exists.
• An unresolved discrepancy around speech output. An unmerged SGLang pull request on GitHub, opened on 27 October 2026 and written by tugot17 (Piotr Mazurek), listed four then-unannounced checkpoints — LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B and LFM3-Omni-2B — and was read as indicating that the audio and omni models only produce text (Orca Router). That conflicts with the released documentation, in which LFM3-Audio-2B is said to produce speech and LFM3-Omni-2B to respond in text or speech. The PR described development checkpoints of unannounced models, so the discrepancy is not necessarily a contradiction — but it is not explained in the launch coverage.
• The vision model's parameter count varies between sources (1.92 billion at MarkTechPost, 1.9 billion at Orca Router). The exact value should be confirmed against the model card.

Why it matters

If the numbers hold, LFM3 moves the boundary of what can run locally on consumer hardware: a text model with 256K context in roughly one gigabyte of memory at 4-bit, with independently unverified claims of beating a 9-billion-parameter model on text benchmarks. For developers of mobile and edge applications — assistants that read documents, listen and answer directly on the device — the combination of a small memory footprint, long context and fast first-token response is the real news. But the $10 million license cap and the absence of independent benchmarks mean anyone considering the model for production should run their own evaluation before committing to it.