Liquid AI’s LFM3 models target local agents, voice and vision

Where the models look strongest

Liquid says its 1.7-billion-parameter text model outperforms its earlier 2.6-billion-parameter version, despite having about 35% fewer parameters. It was pretrained on 22 trillion tokens and then further trained for tool use, business processes and other agent tasks.

The clearest case for testing it is browser automation. In Liquid’s WebArena-Verified results, LFM3-1.7B scored 38.42, compared with 28.94 for Qwen3.5-9B and 3.02 for Liquid’s LFM2.5-2.6B. In AutomationBench, the scores were 41.29, 15.10 and 0.94, respectively. These are company-reported results, not an independent verification.

That strength does not make LFM3 a general replacement for larger models. It trails Qwen3.5-9B on several coding, knowledge and long-context evaluations, and Liquid’s own documentation advises against using it for agent coding or complex, multi-step reasoning because of its size. Liquid also did not directly answer whether the comparisons used identical tools and reasoning capabilities.

A practical pilot should test the whole workflow, not just benchmark scores. For a browser agent, that means measuring tool-call failures, retries and how often work must be handed to another model. Mercedes-Benz and Shopify are named as companies that tested and evaluated the models before launch; the announcement does not provide independent benchmark confirmation or measured customer results.

Smaller models still need a memory budget

LFM3-1.7B supports a 256K-token context window for local use, while Liquid’s cloud console lists a 128,000-token limit. The model itself occupies about 1 GB at four-bit precision, according to Liquid, but that is not the memory budget for a complete application. A separate 1.6 GB figure refers to the key-value cache for a 128K-token context; Liquid has not specified the full memory requirement at 256K.

Liquid’s local speed comparisons cover an AMD Ryzen AI Max+ 395, an Apple M5 Max and an iPhone 17 Pro. The company says LFM3 generated tokens 1.8–2.8 times faster than the next-fastest model in each comparison. The tests used llama.cpp, Q4_0 four-bit models, a 32K-token input and one request at a time. Those conditions matter: the results do not guarantee the same performance on another device or in an application processing audio or video.

For server use, Liquid reports throughput equivalent to 3.6 billion tokens per day on one Nvidia B300 with 1,024 concurrent requests. That is aggregate throughput, not the time one user waits for a complex task. Its latency tests also used different serving software for Liquid’s models and competitors’ models, so the results reflect the model and serving system together—not architecture alone.

The family extends beyond text agents:

For vision, Liquid reports an average score of 71.4 across 62 evaluations for LFM3-VL-2B, compared with 57.6 for LFM2.5-VL-3B and 64.0 for Qwen3.5-2B. On image-text recognition and chart analysis, LFM3-VL-2B scored 83.9, versus 71.2 and 79.1. These results may justify a document-processing pilot, but they do not establish accuracy on any particular company’s documents.

The audio model can analyze recordings and call software tools in response to spoken requests. Liquid says its new encoder processes incoming speech while a person is still speaking, allowing the first response token to appear as soon as 50 milliseconds after the utterance ends. That is the company’s best result, not a measure of full response time or a guarantee for API users.

LFM3-Omni-2B combines vision, speech and language in one configuration and has 2.2 billion parameters. It can analyze video alongside its audio track or take a spoken instruction and an image to produce a tool call. Liquid cites Mercedes-Benz’s exploration of local AI for its MBUX virtual assistant as a possible business scenario, not a confirmed deployment.

The business case depends on the job

I think the interesting question is not whether LFM3 can run locally; it is whether a company can make the complete workflow cheaper and reliable enough to justify doing so. A small model may help with a bounded task such as classifying a support request or extracting a date from an invoice. But the model still needs an application, integrations with business systems and tests on the target hardware.

Liquid’s fine-tuning example shows both the promise and the gap. The company adapted LFM3-VL-2B using about 3,000 public images of wind turbines, trained it for several hours on an Nvidia A100, then evaluated it on about 1,400 images not used for tuning. Liquid reports that classification accuracy rose from less than 10% to 67%. The F1 score for detecting and locating defects, however, was 0.11, up from 0.003. That is evidence that tuning can improve a narrow task, not that the model is ready for autonomous industrial inspection.

For an initial deployment, I would keep the model in an assistive role: prepare a draft inventory update from a worker’s voice note, extract a label from an equipment photo, or flag a possible mismatch for human review. Speech recognition does not establish that the model can identify a failing motor or a warning sound, and video input does not establish continuous monitoring or reliable event detection. Those capabilities need their own tests.

Cloud access offers a simpler way to start. Liquid’s authorized console lists LFM3-1.7B at $0.08 per million input tokens and $0.20 per million output tokens; cached input tokens cost $0.02 per million. Processing one million uncached input tokens and generating one million output tokens would cost $0.28 with that model. The same volume of text tokens costs $0.60 for the vision model and $1.10 for Audio or Omni, before audio-token charges.

All four cloud LFM3 models are listed with a 128,000-token context window. The console says data retention is zero: Liquid says it uses submitted material to fulfill requests but does not save or log it, or use it for training. That is the company’s stated service policy; the materials reviewed do not specify hosting regions or support guarantees.

Local deployment avoids per-token API charges, but it does not eliminate costs. Liquid’s downloadable weights are subject to the LFM Open License v1.0. Under the published terms, companies with annual revenue above $10 million need a separate agreement for commercial use, and corporate license pricing depends on deployment scale. Hardware, power and maintenance still count, as does the risk that an agent sends information through connected external tools even when inference runs on company equipment.

The launch gives buyers enough reason to test narrow workflows, but not enough to assume savings. Those will depend on whether the model’s performance on a company’s own data offsets the equipment, licensing and human review needed to make it dependable.