Liquid AI's LFM3 Family: Text, Vision, Audio, Omni and Decision Models
Liquid AI shipped a full LFM3 lineup in one day, covering text, vision, audio, omni and decision models. Here's what's actually in the release.
What is the LFM3 family?
LFM3 is Liquid AI’s latest generation of small, efficient models released across multiple modalities at once: text, vision, audio, an omni (multi-modal) variant, and a decision model built for agentic and tool-calling tasks. Rather than one flagship release, Liquid pushed out several specialized models on the same day, each aimed at a different slice of the on-device and edge AI market. The flagship entry point most people will test first is the LFM3 1.7B text model, a 1.7 billion parameter model with a hybrid architecture and a 256k context window, small enough to run on a laptop or phone-class device.
TL;DR
- Liquid AI released a full model family in a single day, spanning text, vision, audio, omni and decision variants rather than one flagship model.
- The LFM3 1.7B text model uses a hybrid architecture, supports a 256k context window, and is post-trained for agentic harnesses like Hermes agent.
- At 4-bit quantization the 1.7B model fits in roughly a gigabyte of memory, and it runs in BF16 at around 4.5GB of VRAM, making it viable for consumer hardware.
- Benchmark numbers for the 1.7B model look strong against much larger models, including Qwen’s 9B parameter model, particularly on agentic and tool-use tests.
- Independent hands-on testing found a split result: the model handled structured JSON output, strict instruction following, grounded Q&A, tool calling and multi-step math well, but failed on agentic code-fixing, basic web app debugging, and multilingual translation.
- Coding is the clearest weak spot: the model missed an undefined variable bug and produced a blank HTML page twice, even after being told the output was broken.
- Liquid frames these as specialized models rather than general-purpose flagships, which matters when judging where the benchmarks hold up and where they don’t.
Why release five model types in one day?
Bulk releases are unusual in a field where most labs ship one model at a time and spend weeks on a rollout. Liquid AI’s bet with LFM3 is that specialization beats generalization at small scale. Instead of one model trying to do text, vision, audio and decision-making adequately, the family splits those jobs into separate models tuned for their specific input type and use case. A vision model doesn’t carry the same training burden as a model meant purely for tool calling and agentic workflows. In theory, this lets each model be smaller, faster and cheaper to run while still being competitive on the benchmarks that matter for its niche.
The risk with this approach is obvious: shipping five or more models at once leaves less room to catch quality issues in any single one. That tension shows up clearly once you move from benchmark charts to actual testing, as this article covers below.
How does the LFM3 1.7B text model perform in practice?
On paper, the numbers are striking for a model this small. Liquid’s own benchmark comparisons put LFM3 1.7B against models from Qwen, Google, and MiniCPM, including Qwen’s 9 billion parameter model, a model several times its size. On agent and tool-use benchmarks, the 1.7B model leads or stays competitive with those larger models in most categories. It reportedly trails on agentic coding specifically, which the benchmarks themselves flag as a weaker area.
Hands-on testing backs up a split version of that story. Where the model did well:
- Structured output: asked to extract clean JSON from messy text, it returned valid JSON with correct values and correctly excluded an out-of-stock item, matching the task constraints exactly.
- Strict instruction following: given a task requiring exactly four sentences, no commas, a specific starting letter, and a 60-word limit, it hit every constraint, landing at 43 words.
- Grounded answering: when required to answer only from a provided passage and say “not in the text” when information was missing, it got all three test questions right without fabricating anything.
- Multi-step reasoning: a yield-based math problem (how much raw meat to buy to serve a target plated weight across 150 orders) came back with the correct figure, 71.74 kg, quickly and accurately.
- Tool calling: given a prompt requiring two separate tool calls with different arguments (setting a reminder and checking weather in another city), it returned valid JSON with both tools and arguments correct, with only a minor timezone ambiguity since none was specified in the prompt.
- Hallucination resistance: asked to summarize a non-existent academic paper, it didn’t invent fake findings, though it also didn’t clearly state the paper might not exist.
Where it struggled:
- Agentic debugging: given a broken restaurant inventory dashboard with a planted data bug, the model claimed it had fixed the numbers, but reloading the actual app showed nothing had changed. The underlying values were untouched.
- Code generation: asked to build a self-contained HTML simulation, it produced a blank page twice in a row, even after being told explicitly that the output was blank. A quick code review found an undefined variable, a basic error that undercut confidence in its coding reliability.
- Multilingual translation: asked to translate a simple phrase into around 80 languages, many outputs used incorrect words, wrong scripts, or were left partially untranslated. The model does not appear to be built for broad multilingual use.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Is the LFM3 1.7B model worth running locally?
For a model this size, roughly a gigabyte of memory at 4-bit quantization, the practical appeal is clear: it fits on modest hardware, runs fast, and handles a real cluster of agentic primitives well, structured output, tool calling, instruction following, and grounded Q&A. Those are exactly the building blocks most local agent workflows depend on, which is likely why Liquid post-trained it for harnesses like Hermes agent in the first place.
Where it falls short is anything resembling general-purpose coding or multilingual work. The model card itself reportedly doesn’t claim to be a coding specialist, and the benchmarks show a gap versus larger models in agentic coding specifically. Expecting it to debug a live application or write working front-end code from scratch is asking it to do a job outside its design brief. Used instead as a lightweight backend for tool-calling, JSON extraction, or constrained text generation, it performs closer to its benchmark claims.
What about the vision, audio, omni and decision models?
The text model is the easiest to test quickly because it runs through standard local tooling like llama.cpp, but it’s only one piece of the release. Liquid also shipped a vision model, an audio model, an omni model designed to handle multiple modalities together, and a decision model aimed squarely at agentic and tool-use scenarios. Each is positioned as a specialized tool rather than a do-everything model, following the same philosophy as the text variant: smaller, faster, and tuned for a narrower set of tasks rather than competing head-on with large multimodal flagships.
Because this is a same-day bulk release, independent testing of the other modalities (vision, audio, omni, decision) is still catching up. The pattern seen in the text model, strong benchmark claims paired with mixed real-world results depending on the task, is a reasonable baseline expectation until each sibling model gets the same kind of hands-on scrutiny.
Frequently Asked Questions
What is LFM3?
LFM3 is Liquid AI’s newest family of models released across text, vision, audio, omni and decision variants, emphasizing small size and efficiency over raw general-purpose scale.
How big is the LFM3 1.7B model and what hardware does it need?
It has 1.7 billion parameters, uses a hybrid architecture, and supports a 256k context window. At 4-bit quantization it fits in roughly a gigabyte of memory; running in BF16 consumes around 4.5GB of VRAM.
Is LFM3 1.7B good at coding?
Not reliably, based on hands-on testing. It failed to fix a planted bug in a web app and produced broken HTML code twice in a row, missing a basic undefined-variable error. Liquid’s own benchmarks also show it trailing larger models specifically on agentic coding.
What is LFM3 1.7B actually good at?
Testing found it performed well on structured JSON output, strict instruction following, grounded question answering, multi-step math reasoning, and tool calling with correct argument filling, tasks that align closely with agentic harness use cases.
Can LFM3 1.7B handle multiple languages?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Testing across roughly 80 languages found frequent errors, including invented words, wrong scripts, and partially untranslated output, suggesting it is not designed as a general multilingual translation model.
