Liquid AI's LFM3 Shrinks Six Specialized Models to Run on Your Phone

Liquid AI has released LFM3, a family of six small foundation models designed for single-GPU fine-tuning and deployment on phones, laptops, edge systems, and cloud hardware. Each variant branches from a 1.7-billion-parameter backbone and targets text, vision, audio, multimodal input, or structured decision-making. The weights are available through Hugging Face, with hosted access through Liquid and partner platforms.

For developers, the release combines several practical constraints in one family: models near 2 billion parameters, a context window of up to 256K tokens, support in common inference runtimes, and modality-specific variants that can be adapted without training a larger general-purpose model. Liquid says the 1.7B text model outperforms its previous 2.6B model with 35% fewer parameters.

Benchmark, latency, throughput, and robotics results cited here come from Liquid’s release materials. Cross-model comparisons can vary with prompts, quantization, hardware, and runtime configuration, so deployment teams should reproduce the relevant tests on their own workloads.

Hardware steers the architecture

LFM2 combined short convolutions with attention to accelerate on-device inference. LFM3 adds a third operator type and uses automated architecture search to balance model quality with performance on specific hardware.

Liquid expanded its search from the phone and laptop CPUs used for LFM2 to 14 device and backend combinations. The targets included the Samsung Galaxy S26 Ultra, OnePlus 15, Raspberry Pi 5, Apple M4 Max, AMD Ryzen AI hardware, Arm and x86 cloud CPUs, GPUs, and Qualcomm’s Hexagon neural processing unit.

The 24-layer backbone contains three kinds of computation:

• 13 gated short-convolution layers capture local patterns with a kernel size of seven, increased from three in LFM2.
• Five sliding-window attention layers attend to the most recent 256 tokens, limiting the cost of local context processing.
• Six global-attention layers connect information across the full available context.

Head-wise gating applies a sigmoid gate to each attention head, allowing the model to suppress heads that contribute little. Liquid says this also reduces activation outliers that can degrade low-bit quantization. Global-attention layers use partial rotary position embeddings, or partial RoPE, which applies position rotation to the high-frequency half of each head while preserving content matching in the low-frequency dimensions.

Liquid reports that LFM3-1.7B matches LFM2-1.2B in its per-parameter short-prompt comparison and runs about 25% faster with an 8K-token context. The base model was pretrained on 22 trillion tokens. Its final 3.5 trillion tokens formed a higher-quality mid-training stage rich in tool calls and agent traces, which are examples of models selecting tools and completing multistep tasks. That stage also extended the supported context to 256K tokens.

Six branches cover the deployment map

Each public model begins with LFM3-1.7B-Base, then receives modality-specific components and post-training for its intended workload.

The text model targets local agents

LFM3-1.7B received four post-training stages. Liquid used 1.75 trillion supervised fine-tuning tokens across chat, reasoning, tool calling, and agent traces, followed by mixed-domain reinforcement learning with a length penalty. The company then distilled behavior from six specialist reinforcement-learning teachers and ran a final reinforcement-learning stage inside real agent harnesses, the software environments that expose tools, state, and task feedback to a model.

On Liquid’s agent evaluations, LFM3-1.7B scored 45.6 on LiveClaw and 38.42 on WebArena-Verified. Qwen3.5-9B scored 35.77 and 28.94, respectively. LFM3 also reached 71.39 on BFCL V4, which evaluates function-calling behavior.

Long-context memory use is a central deployment advantage. A 128K-token context requires 1.6 GB of key-value cache, the stored attention state used during generation. Liquid measured 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B under its comparison setup. With llama.cpp and Q4_0 four-bit quantization, LFM3-1.7B decoded 1.8 to 2.8 times faster than the next model Liquid tested across three target devices.

The vision encoder uses fewer FLOPs

LFM3-VL-2B includes a new 164-million-parameter encoder that combines convolution and attention. Liquid says it uses about one-third of the compute required by the SigLIP2 encoder in earlier LFM vision-language models. It averaged 55.9% accuracy across 13 encoder benchmarks, matching the reported result for Qwen3.5-2B’s vision encoder with 2.2 times fewer floating-point operations.

For video input, the model embeds pairs of sampled frames into one temporal slice, cutting the resulting token sequence in half. Shorter visual sequences reduce attention work and leave more of the context window available for instructions, conversation history, or additional media.

Across 62 vision benchmarks, Liquid reports a lead of at least 7.4 points over similarly sized models. The 2B model matched Gemma 4 12B, exceeded Gemini 3.5 Flash Lite, and finished within two points of GPT-5.6 Luna in the company’s aggregate comparison. Agentic vision remained a weaker area, with a 42.5 average and a substantial gap to frontier systems.

Audio processing begins mid-utterance

LFM3-Audio-2B uses a fully causal encoder built from sliding-window attention and short convolutions. It can process incoming audio while a speaker is still talking, reducing the work required after the utterance ends. Liquid measured time to first token as low as 50 milliseconds, though complete application latency will also depend on endpoint detection, networking, tool execution, and audio playback.

The accompanying LiquidCodec tokenizer compresses 80 milliseconds of speech into 12 discrete tokens that the model can predict in parallel. On one CPU thread of an Apple M4 Max, Liquid measured speech decoding at about 87 times real time.

The audio model scored 58.5 on Liquid’s internal tool-calling evaluation, compared with 44.3 for Qwen2.5-Omni-3B. Its reported English word error rate was 6.2% for automatic speech recognition. General audio understanding remained less competitive: the model scored 50.6 on MMAR, while MiMo-Audio-7B scored 62.3.

Decision models return schemas and confidence

The d1 models accept an input and a list of questions, then answer them together in one forward pass. Their outputs follow defined types and include calibrated probabilities, which are confidence estimates trained to track observed accuracy. This structure can reduce the parsing and validation required when routing requests, extracting fields, or applying policy rules.

The typed-output system supports four operations:

1. Point: locate every matching object with bounding boxes and counts.
2. Read: extract text, numbers, or dates with source spans, including a not-found result when the value is absent.
3. Estimate: return a numerical value with a likely range.
4. Pick: select every applicable option and assign each one a probability.

Liquid says d1-2B reduces the parameter count of d1-3B by one-third while improving every benchmark in its tracking suite. It scored 63.2 on the company’s 117-benchmark meta-index, up from 58.8 for d1-3B. The d1-Omni-2B variant adds audio input, allowing the same structured routing system to process spoken and typed requests.

One A100, 3,000 inspection photos

Liquid demonstrates its specialization strategy with a wind-turbine inspection task. Starting with about 3,000 public images, the company applied supervised fine-tuning and Group Relative Policy Optimization, a reinforcement-learning method commonly abbreviated as GRPO. A few hours of training on one Nvidia A100 raised LFM3-VL-2B’s defect-classification accuracy from below 10% to 67%.

Detection F1, a metric that combines precision and recall, increased from 0.003 to 0.11. GPT-6 Luna scored 0.01 out of the box. Because LFM3 received task-specific training and the GPT model did not, the comparison measures the complete deployment setup as well as underlying model capability.

Liquid also tested LFM3-VL-2B as a vision-language-action model for robotics, using parallel prediction to generate chunks of control actions. It reported 96.2% average success across four LIBERO benchmark suites and 45.7% on RoboCasa-GR1 tabletop tasks. On a Jetson AGX Orin, TensorRT with 16-bit floating-point inference generated action chunks in under 50 milliseconds. End-to-end control latency will also include sensing, planning integration, and actuation.

Start with the deployment constraint

Developers can download the weights from Hugging Face or use Liquid’s hosted API and Playground. The launch also includes access through OpenRouter, Qualcomm AI Hub, and Vercel, plus runtime support for llama.cpp, MLX, vLLM, Ollama, LM Studio, Unsloth, and SGLang. Pipette provides comparisons with other on-device models.

A practical evaluation should follow the application’s real constraints:

1. Select the modality: choose text, vision, audio, omni, or typed decisions based on the production input and output contract.
2. Check the model card: confirm licensing, permitted uses, prompt templates, preprocessing requirements, and context limits for the selected variant.
3. Benchmark the target device: test the expected context length, batch size, quantization format, and runtime on production-class hardware.
4. Fine-tune against held-out data: measure task accuracy before and after adaptation, using examples that reflect production failures and edge cases.
5. Profile the complete pipeline: include media preprocessing, retrieval, tool calls, networking, validation, and output decoding in latency and memory measurements.

LFM3 gives teams a common small-model base with six deployment paths. Its practical value will depend on whether a selected branch preserves enough task accuracy after quantization and meets the memory, latency, and licensing requirements of the target system.