# Liquid AI LFM3 Puts Open-Weight 2B Models at the Center of On-Device AI

> The six-model LFM3 family runs text, vision, audio, omni, and decision workloads on a shared 1.7B backbone, targeting phones, laptops, and edge hardware over cloud-scale deployment.

*Published 2026-10-29 · By Marcus Vance*

LFM3 is Liquid AI's third-generation family of open-weight foundation models — six 2B-class models spanning text, vision, audio, omni, and decision modalities, all branched from a single LFM3-1.7B-Base backbone with 1.73 billion parameters.

Liquid AI announced and released the LFM3 family on Oct. 27, 2026, marking a deliberate turn toward specialized, on-device intelligence at a moment when much of the industry is scaling frontier models into the cloud. The six open-weight models — text, vision, audio, omni, and two decision variants — all descend from the same LFM3-1.7B-Base checkpoint, a 1.73 billion-parameter backbone pre-trained on 22 trillion tokens. By concentrating compute on a shared backbone rather than proliferating dense models, Liquid AI is betting that most production AI workloads are better served by small, task-specific systems that fit in a gigabyte of memory.

The release is the company's clearest statement yet of its specialized-intelligence thesis. Rather than chase parameter counts against competitors, Liquid AI engineered LFM3 to run where the data lives: on phones, laptops, and edge hardware. The text model alone supports a 256K-token context window and, at 4-bit quantization, occupies approximately 1 GB of memory, a footprint that allows it to execute locally on consumer hardware. That positioning also shapes the business model: open weights under the LFM Open License v1.0, single-GPU fine-tuning, and partnerships with silicon vendors such as AMD, Qualcomm, and Ambarella.

## How did Liquid AI get to a specialized 2B-class family?

Liquid AI was founded by researchers with roots in MIT's CSAIL, and its earlier LFM generations — LFM1 and LFM2.5 — established the company's signature tri-hybrid architecture, which mixes gated convolutions with attention mechanisms instead of relying on dense transformer stacks. LFM3 is the third generation of that design and the first to organize an entire product family around a single backbone. The company describes the move as building for “specialized intelligence,” a term it uses to distinguish narrow, deployable capability from the broad-but-static competence of cloud-scale models.

The strategic shift is visible in the training numbers. LFM3's base model was pre-trained on 22 trillion tokens, with the final 3.5 trillion tokens dedicated to a mid-training stage focused on agentic traces, tool calls, and context extension to 256K tokens, according to Liquid AI. That training curriculum is explicitly designed for agents that must call tools and hold long conversations, rather than for open-ended chat. The result, per Liquid AI's published benchmarks, is a model that outperforms far larger rivals on general knowledge and agentic task completion alike.

## What does the LFM3 family include and how do the six models differ?

All six LFM3 models branch from the LFM3-1.7B-Base checkpoint, but each receives different post-training to serve a distinct modality. The lineup spans the input types that production systems typically need: text, image, video, audio, and combined streams, plus two decision models designed for structured output.

*The six LFM3 models, their modalities, and key specifications, plus the shared base checkpoint*

| Model | Modality | Key specifications |
| --- | --- | --- |
| LFM3-1.7B-Base | Backbone | 1.73B parameters; pre-trained on 22T tokens; 256K-token context |
| LFM3-1.7B | Text | 256K-token context; roughly 1 GB at 4-bit quantization |
| LFM3-VL-2B | Vision-language | Images and video understanding |
| LFM3-Audio-2B | Audio | Speech and audio input |
| LFM3-Omni-2B | Omni | Combined text, vision, and audio |
| d1-2B | Decision | Typed answers with calibrated probabilities |
| d1-Omni-2B | Decision + omni | Multimodal typed outputs |

The two decision models, d1-2B and d1-Omni-2B, are the most distinctive additions. They return typed answers with calibrated probabilities rather than free-form text, a design choice that matters for automation pipelines where downstream systems need confidence scores, not prose. Calibrated probabilities allow developers to set thresholds for when a model should act autonomously and when it should hand off to a human, a requirement that has become central to agentic deployments in regulated industries.

The vision and audio variants extend the same backbone to non-text inputs, and the omni model combines all modalities in a single checkpoint. Because all six share the base architecture, organizations can fine-tune one model and apply learnings across the family — or deploy multiple variants from a single training pipeline.

## How does LFM3's architecture deliver long context at small scale?

The efficiency story starts with the tri-hybrid architecture. According to the model card on Hugging Face, LFM3-1.7B-Base stacks 24 layers composed of 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers. The gated convolutions handle local pattern extraction cheaply, while the two attention types manage short- and long-range dependencies where they are actually needed. That division of labor is what lets a 1.73 billion-parameter model sustain a 256K-token context window without the quadratic memory costs of dense full attention.

Liquid AI reports the base checkpoint was pre-trained on 22 trillion tokens. The final 3.5 trillion tokens formed a mid-training stage concentrated on agentic traces, tool calls, and context extension — a curriculum that trains the model for the behaviors developers actually deploy, rather than for generic next-token prediction. At 4-bit quantization the text model fits in about 1 GB of memory, which is what makes phone- and laptop-class deployment feasible.

- Pre-training on 22 trillion tokens establishes broad knowledge and linguistic competence in the LFM3-1.7B-Base checkpoint.
- A mid-training stage over the final 3.5 trillion tokens focuses on agentic traces, tool calls, and extending context to 256K tokens.
- Post-training branches create the six specialized variants — text, vision, audio, omni, and two decision models — from the shared base checkpoint.
- Single-GPU fine-tuning adapts any variant to a specific domain or enterprise task without a dedicated training cluster.

## How does LFM3 compare with much larger open-weight models?

Liquid AI's published evaluations position LFM3-1.7B against models several times its size. Across 29 benchmarks, LFM3-1.7B averaged 49.37, compared with 47.24 for Qwen3.5-9B — a model roughly five times larger. The gap is more dramatic on agentic tasks. LFM3-1.7B scored 38.42 on WebArena-Verified, a 12.7x improvement over LFM2.5-2.6B's score of 3.02, while using 35% fewer parameters.

The WebArena-Verified result is the one that matters most for the agentic workloads Liquid AI is courting. WebArena-Verified measures how well a model completes real web tasks — navigating sites, filling forms, and executing multi-step operations — which is precisely the category of work that on-device assistants and automation agents perform. A 12.7x improvement over the prior generation, with fewer parameters, signals that the mid-training curriculum on agentic traces and tool calls paid off in measurable behavior change, not just benchmark cosmetics.

The comparison to Qwen3.5-9B is also notable for what it implies about the open-weight market. If a 1.7 billion-parameter model can match or beat a 9 billion-parameter model on aggregate benchmarks, the marginal value of mid-size cloud-hosted models shrinks for the long tail of tasks that fit in small footprints. The remaining gap is in the long tail of hard reasoning problems, where larger models still hold an edge — but Liquid AI's results suggest that edge is narrower than the parameter counts imply.

## What does LFM3 mean for on-device AI and enterprise adoption?

The release lands alongside an infrastructure push. Liquid AI lists AMD, Qualcomm, and Ambarella as partners, a signal that LFM3 is designed for the silicon that ships in laptops, phones, and edge devices rather than for data-center racks. For enterprises, the appeal is operational: a model that runs locally removes per-token API costs, keeps data on-device, and eliminates the latency of a round trip to the cloud. At roughly 1 GB, the text model fits comfortably alongside an application binary on a modern phone or laptop.

The licensing and fine-tuning story reinforces that pitch. All six models are available as open weights on Hugging Face under the LFM Open License v1.0, and every variant supports single-GPU fine-tuning. That combination lowers the barrier for teams that want to adapt a model to proprietary data without building a training cluster. For a mid-market company, fine-tuning a 2B-class model on one GPU is a materially different cost equation than fine-tuning a 70B model — or paying for API access to one.

There is also a strategic angle for hardware makers. As on-device models get capable enough for real agent work, silicon vendors gain a reason to ship more neural processing capacity. The AMD, Qualcomm, and Ambarella partnerships suggest Liquid AI expects LFM3 to become a reference workload for the next generation of edge processors, in the same way that early LLMs became reference workloads for data-center GPUs.

## What do developers say about running LFM3 on real workloads?

> “Small models are ready for real agent work on the Mac, and LFM3 is the proof. LFM3 is the first sub-2B model we’ve tested that is ready for real agent work on a Mac — it beat a larger model on task success while using 58% fewer tokens and almost never getting stuck.”Ivan Petrukha, Staff AI Engineer, LLM Efficiency Lead at MacPaw

The MacPaw evaluation, from a company that builds consumer software including cleanup and security tools, is one of the first independent signals that LFM3's efficiency claims translate into practice. MacPaw's testing found LFM3 beat a larger model on task success while consuming 58% fewer tokens and “almost never getting stuck” — a qualitative marker that agentic loops, which often fail when a model repeats the same failed action, stayed productive.

Liquid AI's own framing ties the release to a broader product thesis. “Today, we’re excited to announce LFM3, our new generation of Liquid Foundation Models (LFMs) built for specialized intelligence,” the company said. “The LFM3 model family is multimodal by nature and brings text, vision, audio, omni, and decision models onto one new shared backbone.” The word “specialized” is doing a lot of work here: it signals that Liquid AI is not trying to out-scale the largest labs, but to win the workloads where efficiency and modality coverage decide the outcome.

## What comes next for LFM3 and the specialized-model category?

The decision models point to where Liquid AI is headed next. Typed outputs with calibrated probabilities are a prerequisite for reliable automation, and they position LFM3 as a component in agent pipelines rather than a chat endpoint. As agentic frameworks mature, the bottleneck is no longer raw reasoning capability but reliability — knowing when a model is confident enough to act. The d1 models are an explicit bet that calibrated decision-making, not parameter count, will define the next wave of production AI.

The 22 trillion-token pre-training run and the 3.5 trillion-token agentic mid-training stage also hint at the company's roadmap: future generations are likely to push further into tool use and long-horizon tasks while keeping the family structure that lets one backbone serve many modalities. For enterprises, the practical question is whether a 2B-class model can replace a mix of cloud APIs for common workflows. The benchmark data says yes for agentic web tasks and general knowledge; the open question is how the models hold up in production at scale, where edge cases accumulate.

None of this implies the cloud-scale frontier is obsolete. The largest labs continue to push reasoning and multimodal boundaries that small models cannot reach. But LFM3's release, combined with the MacPaw evaluation and hardware partnerships, suggests the market is splitting: frontier models for the hardest problems, and small specialized models for the vast majority of deployed workloads. Liquid AI is staking its position on the latter — and the benchmark numbers say the gap between the two categories is closing faster than the parameter counts suggest.

## Sources

1. [Announcement and release of the six-model LFM3 family on Oct. 27, 2026, sharing a single LFM3-1.7B-Base backbone; pre-training on 22 trillion tokens; 256K-token context window; about 1 GB memory at 4-bit precision; 49.37 average across 29 benchmarks versus Qwen3.5-9B's 47.24; WebArena-Verified score of 38.42 versus LFM2.5-2.6B's 3.02; LFM Open License v1.0; single-GPU fine-tuning; AMD, Qualcomm, Ambarella, and MacPaw involvement](https://www.liquid.ai/blog/lfm3)
2. [LFM3-1.7B-Base specifications: 1.73 billion parameters; pre-trained on 22 trillion tokens; 256K-token context length; about 1 GB memory at 4-bit precision; tri-hybrid architecture with 24 layers comprising 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers](https://huggingface.co/LiquidAI/LFM3-1.7B-Base)
3. [LFM3-1.7B text model specifications: 256K-token context window; about 1 GB memory at 4-bit precision; 1.73 billion parameters; 24-layer tri-hybrid architecture with 13 gated short convolutions, 5 sliding-window attention, and 6 global attention](https://huggingface.co/LiquidAI/LFM3-1.7B)
4. [Liquid AI released LFM3, six open-weight 2B-class models (text, vision, audio, omni, plus two decision models) on a 1.7B backbone. The text model supports 256K context, fits ~1GB at 4-bit quantization, and runs on…](https://liquid.ai)

---
Source: https://aiintelreport.com/frontier-models/liquid-ai-lfm3-open-weight-2b-on-device
Index: https://aiintelreport.com/llms.txt · Full text: https://aiintelreport.com/llms-full.txt