On October 27, 2026, Liquid AI released LFM3, a new family of AI models designed to run on devices. The family consists of six models covering text, image and audio processing as well as decision-making tasks, all built on a shared base model. The text model has roughly 1.7 billion parameters, and according to the company, its weights fit in about 1GB when quantized to 4 bits. But a small model is not suited to every use. Supported input formats, context lengths and output methods differ from model to model, so companies adopting LFM3 for document processing, voice control or similar work need to look beyond published model sizes and assess real-world accuracy and the runtime environment.
Six models from one base, with different context lengths and input/output formats
All six LFM3 models are built on a common base, "LFM3-1.7B-Base." The family includes "LFM3-1.7B" for text, "LFM3-VL-2B" for image and video understanding, "LFM3-Audio-2B" for speech, and "LFM3-Omni-2B," which integrates text, images and audio. Two decision-focused models, "d1-2B" and "d1-Omni-2B," round out the lineup. The base model itself is also public, bringing the total number of available checkpoints to seven.
LFM3-VL-2B is the first Liquid AI vision model to support video input. LFM3-Omni-2B is the company's first model to integrate images, audio and language. The d1 decision series has also expanded. In addition to binary judgments and scoring, it can now return structured results such as information extracted from documents, positions within images and numeric ranges. d1-Omni-2B extends these decision capabilities to audio input.
The internal architecture has changed from the previous generation. LFM2 combined short convolutions with attention, while LFM3 adds local attention, giving it three types of processing. Of the 24 layers, 13 are short convolutions, 5 are local attention and 6 are global attention. The local attention window is set to 256 tokens.
According to Liquid AI, LFM3 was pretrained on a total of 22 trillion tokens. In the final 3.5 trillion tokens, the company increased data on AI agent execution logs and tool calls, and extended the context length to 256K tokens. However, the published information does not isolate how much of the performance gain comes from the architecture changes alone. Model weights are public, and the models are also available through Liquid AI's Playground and API.
When choosing a model in practice, an important factor is the context length it can handle at one time. Context length is the upper limit on combined input and output tokens, and it varies by model even within the LFM3 family. Comparing the specifications on the model cards published on October 27, 2026 gives the following:
| モデル | 対応する入力 | 出力 | 文脈長（入力＋出力） |
|---|---|---|---|
| LFM3-1.7B | 文章 | 文章 | 256K |
| LFM3-VL-2B | 文章・画像・動画 | 文章 | 128K |
| LFM3-Audio-2B | 文章・音声 | 文章・音声 | 32K |
| LFM3-Omni-2B | 文章・画像・音声・動画 | 文章・音声 | 32K |
The text model has a 256K-token context length, while the image and video model is limited to 128K and the audio and Omni models to 32K. In other words, even models developed from a common base do not handle the same amount of information or offer the same input and output capabilities. This table compares published specifications; it is not the result of measuring processing speed or business accuracy on the same hardware. Also, being able to accept a long document is not the same as understanding all of it and answering accurately.
In addition to reading documents, charts and screens, LFM3-VL-2B can ingest video at 2 frames per second. LFM3-Audio-2B supports 18 languages, including Japanese, and also supports tool calling by voice. Liquid AI says that because the model begins processing input while the user is still speaking, the time to the first output token after the user finishes can be as short as 50 milliseconds. However, this figure is not the time to complete an entire spoken response, nor does it guarantee the same responsiveness in every runtime environment.
LFM3-Omni-2B can handle multiple formats, such as images and audio, in a single model. Liquid AI, however, positions it as a "non-reasoning model," and it does not support enabling a reasoning mode. Available features also differ by runtime format. The SGLang version, which runs at BF16 precision, supports text output, while the llama.cpp version, which uses the GGUF format, also comes with a real-time audio server. Even for the same model, you need to check individually whether the runtime you plan to deploy supports the audio features you need.
Decision-focused "d1" model: watch for accuracy drop on long inputs
LFM3's decision model "d1-2B" is not a model that generates text sequentially like a typical chat AI. Given text, images or video, it outputs typed answers and probabilities for multiple questions in a single forward pass. This makes it suited to tasks that require retrieving information in a fixed format, such as extracting dates, amounts and relevant locations from an invoice. However, although d1-2B's nominal context length is 128K tokens, its official model card states that answer reliability declines beyond roughly 64K tokens. Being able to accept up to 128K tokens does not mean it maintains equal accuracy across that entire range.
Beats Qwen3.5-9B at browser tasks, but trails on coding and long-text processing
In benchmark results published by Liquid AI, the text model LFM3-1.7B outperformed the much larger Qwen3.5-9B on some AI agent evaluations. It did particularly well on browser operation and function calling, while Qwen3.5-9B recorded higher scores on complex tool integration, coding and long-text processing. The results for both models are as follows:
| ベンチマーク | LFM3-1.7B | Qwen3.5-9B |
|---|---|---|
| WebArena-Verified | 38.42 | 28.94 |
| BFCL V4 | 71.39 | 62.83 |
| MCP-Atlas | 32.57 | 42.67 |
| SWE-bench Verified | 43.40 | 51.20 |
| RULER 128K | 65.56 | 87.11 |
On WebArena-Verified, which measures browser operation ability, LFM3-1.7B scored 38.42, ahead of Qwen3.5-9B's 28.94. On BFCL V4, which evaluates function-calling ability, the smaller LFM3 also led, 71.39 to 62.83. On the other hand, Qwen3.5-9B came out ahead on MCP-Atlas, which evaluates tool integration using MCP (32.57 vs. 42.67), and on SWE-bench Verified using OpenCode (43.40 vs. 51.20). On RULER, which measures 128K-token long-text processing, LFM3-1.7B scored 65.56 against Qwen3.5-9B's 87.11. Strong results in specific browser and function-calling tasks do not necessarily mean strong performance in complex agent work, coding or long-document comprehension.
All of these scores were measured by Liquid AI and are not independent third-party evaluations. Measurements used vLLM and BF16 precision, with the maximum context length for both LFM3 and Qwen set to 262,144 tokens. Each benchmark also has its own conditions, such as the number of execution steps, which should be taken into account when comparing results. At the very least, the results do not show that the same scores would be obtained when running 4-bit quantized LFM3 on an actual device.
The official LFM3-1.7B model card itself acknowledges the limits of a small model. In particular, it states that the model is not recommended for agentic coding or complex multi-step reasoning. Even though it performs well on browser operation and relatively simple tool calls, those results should not be applied to advanced autonomous agents in general.
The roughly 1GB figure is also a guideline for the text model quantized to 4-bit precision. It does not guarantee total memory usage, which includes the memory used by the whole application, the working space needed for long inputs, and the KV cache. Whether the model runs smoothly on a real device needs to be measured after settling on the model format, input length and runtime environment.
Wind turbine blade damage detection improves with fine-tuning, but localization remains a challenge
Liquid AI says all LFM3 models can be fine-tuned on a single GPU. As an example, it published an experiment on detecting wind turbine blade damage from images. The company used about 3,000 images from the public dataset "WTBs2025" and ran several hours of additional training on a single NVIDIA A100 GPU, combining supervised fine-tuning with GRPO (Group Relative Policy Optimization).
When the fine-tuned model was evaluated on about 1,400 images not used in training, classification accuracy was 67%. The F1 score, which measures how well damage locations are detected, improved from 0.003 before tuning to 0.11, according to the report. The experiment shows that a model can be adapted with domain-specific data even with limited compute, but the numbers need careful interpretation.
Classification accuracy and localization F1 measure different abilities. Better classification of whether an image contains damage does not mean the model can accurately pinpoint where the damage is. In fact, the localization F1 remains at 0.11 even after improvement, making it hard to conclude that practically sufficient detection performance was achieved. The result is a limited experiment by Liquid AI and does not constitute independent verification or proof of successful field deployment. Whether fine-tuning on specialized data is effective must be judged against the accuracy requirements and operating conditions of the deploying organization.
Beyond on-device performance, check commercial licensing and data transmission paths
LFM3 model weights are public, but that does not mean unconditional commercial use. The "LFM Open License v1.0," which applies to the text model, includes a revenue condition for commercial use, with the threshold set at US$10 million in annual revenue. However, the license text contains differing wording about that boundary, so businesses with revenue near the threshold need to confirm the specific conditions that apply.
The license's definition of an entity also includes companies controlled by, controlling, or under common control with that entity. As a result, usage conditions cannot necessarily be judged from a subsidiary's revenue alone. For corporate groups, it is important to confirm the scope of the entities covered.
Data handling also needs attention. Even when the model itself runs on a device such as a smartphone or PC, data may be sent to external destinations if the application calls external web services or tools. Running AI locally is not the same as having all processing stay on the device. For work involving confidential information, you need to check not only where the model runs but also integrations with external tools and data transmission paths.
By deploying purpose-specific models from a shared base, the LFM3 family broadens the options for small AI. But text processing, image understanding, voice control and decision tasks each demand different capabilities, and public benchmarks alone cannot determine the best model. When adopting LFM3, first clarify the required input and output formats, then measure accuracy and response speed with real business data. After that, checking device memory requirements, feature differences across runtime formats, external data transmission paths and commercial license terms makes it possible to judge realistically which tasks can be entrusted to a small model.