LFM3

LFM3 is the third generation of Liquid Foundation Models (LFMs), a family of open-weight small language models and multimodal models released by Liquid AI on October 27, 2026.^[1][2] The first release contains six post-trained models of about 2 billion parameters each, covering text, vision, audio, combined "omni" input, and two decision models, plus the pre-trained base checkpoint they were all built from.^[1][3][4] Every model shares one backbone, LFM3-1.7B-Base, which uses what Liquid AI calls a "tri-hybrid" architecture of gated short convolutions, sliding-window attention, and global attention.^[1][4] Liquid AI positions the family for what it calls "specialized intelligence": models small enough to fine-tune on a single GPU and run on phones, laptops, and other edge devices.^[1][2]

Release and availability

Liquid AI announced LFM3 in a blog post titled "Introducing LFM3: Every Modality for Specialized Intelligence" and in a thread on X on October 27, 2026.^[1][2] The company said the base and post-trained models were open-weight and available the same day on Hugging Face, in its Playground, and through the Liquid AI API.^[1] On Hugging Face the models are grouped in an LFM3 collection that the model cards describe as including "base, text, vision, audio, omni, and decision models."^[3]

The six post-trained models are:^[1][2]

Liquid AI also published the pre-trained LFM3-1.7B-Base checkpoint, which it recommends only for heavy customization such as continued pre-training, language- or domain-specific assistants, or new post-training methods.^[4] Alongside the BF16 weights, the company released quantized and converted checkpoints in GGUF, MLX, ONNX, FP8, and NVFP4 formats for runtimes such as llama.cpp, Ollama, LM Studio, MLX, ONNX Runtime, vLLM, and SGLang; the exact set of formats differs from model to model.^[3][5][7]

The text, vision, and decision models list support for 16 text languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish.^[3][5][8]

Architecture

Tri-hybrid backbone

The earlier LFM2 generation, described on the Liquid AI page, was a hybrid of short convolutions and attention found by an automated, hardware-in-the-loop architecture search.^[1] Liquid AI says that for LFM3 it used the same search to redesign the architecture for agentic use cases, which need more capability per parameter and fast inference over long contexts.^[1] It also widened the hardware in the loop, from CPUs in one phone and one laptop for LFM2 to 14 device and backend combinations for LFM3: phones (Samsung Galaxy S26 Ultra and OnePlus 15), a Raspberry Pi 5, laptops (Apple M4 Max and AMD Ryzen AI), Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU.^[1] Liquid AI's launch post on X described this as optimization "for 14 processor classes across GPUs, CPUs, and NPUs."^[2]

According to Liquid AI, gated short convolutions still make up more than half of the layers. To add capability without slowing long-context inference, the team tried swapping some of them for other sequence blocks and found that sliding window attention (SWA) gave the best balance of quality and speed.^[1] The result is the layer mix Liquid AI calls tri-hybrid:^[1][3]

Liquid AI lists three further changes, which it says it kept only where they improved quality without a meaningful speed cost:^[1]

• Head-wise gated attention. A sigmoid gate scales each attention head's output so a head can switch off when it has nothing to add. Liquid AI says this can help suppress activation outliers that hurt quantization and can improve training stability.
• Partial RoPE. Global attention layers apply rotary position embedding only to the highest-frequency half of each head, leaving the low-frequency half unrotated so that those dimensions match on content regardless of distance. SWA layers keep full RoPE.
• Sandwich norm. Each block normalizes both its input and its output.

LFM3 uses the same 128,000-token vocabulary as LFM2.5-8B-A1B, which Liquid AI says was built to tokenize non-Latin scripts more efficiently.^[1][3] Liquid AI says the payoff shows at long context: per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices and is about 25% faster with 8K tokens in context.^[1]

Multimodal encoders

LFM3-VL-2B uses a new in-house vision encoder instead of the SigLIP2 encoder used in earlier LFM vision-language models.^[1] The model card describes it as 14 layers (8 gated depthwise 2D convolutions and 6 bidirectional attention layers) operating on 16x16 patches with 2D rotary embeddings; images and video are processed at native resolution, each 2x2 block of patches becomes one language token, and video is sampled at 2 frames per second.^[5] Liquid AI says the 164M-parameter encoder uses three times less compute than SigLIP2 and matches the vision encoder of Qwen3.5-2B at 55.9% average accuracy across 13 vision-language benchmarks while needing 2.2 times fewer FLOPs.^[1] For video, the encoder embeds each pair of sampled frames into one temporal slice of visual features, halving the temporal sequence length.^[1]

The audio models use a new causal audio encoder built from sliding-window attention and short-convolution layers, which Liquid AI says keeps memory use bounded and lets the model prefill while a user is still speaking; it reports time to first token as low as 50 ms.^[1][6] Speech output comes from LiquidCodec, a new audio tokenizer built on the LFM3 architecture. It represents each 80 ms of speech with 12 discrete tokens, supports streaming decoding, and lets the model predict several audio tokens in parallel rather than one at a time as in previous LFM audio models. Liquid AI says LiquidCodec converts tokens to speech at about 87 times real time on a single CPU thread of an Apple M4 Max.^[1][6]

Training

LFM3-1.7B-Base was pre-trained on 22 trillion tokens in two stages.^[1][4] Compared with LFM2, the main pre-training mix adds more scientific papers, repository-level code, math, and a wider range of knowledge sources.^[1][4] The final 3.5 trillion tokens form a mid-training stage on curated data with a larger share of agentic traces and tool calls; this stage also extends the context window to 256K tokens.^[1][4]

Post-training of the text model LFM3-1.7B had four stages, according to Liquid AI:^[1]

LFM3-VL-2B starts from the same base model. Its post-training adds a vision mid-training stage that joins the base model with the new vision encoder, followed by supervised fine-tuning, vision reinforcement learning, and multi-domain on-policy distillation from a vision teacher and the mixed-RL text model.^[5]

Models

LFM3-1.7B (text)

LFM3-1.7B is a text model "for local agents and specialized applications." Liquid AI says it is post-trained to run in agent harnesses such as Hermes Agent, Pi, and OpenClaw and takes about 1 GB of memory at 4-bit precision.^[1][3] The company claims it outperforms its previous LFM2.5-2.6B while using 35% fewer parameters.^[1] The model card recommends it for local agents, tool calling, structured outputs, and domain-specific fine-tuning, advises using retrieval or search for knowledge-heavy tasks, and says it is not recommended for agentic coding or complex multi-step reasoning because of its size.^[3]

Liquid AI published results on 29 benchmarks across agents, tool use, agentic coding, coding, instruction following, knowledge, and long context. Selected figures, all reported by Liquid AI:^[1][3]

Liquid AI summarized these results as LFM3-1.7B competing with models several times its size on agentic tasks, leading on instruction following and on function calling in BFCL and ToolSandbox, and beating Qwen3.5-9B on long tasks in browsers and business apps, while acknowledging that larger models still win on agentic coding and knowledge.^[1] It added that access to search nearly doubles the model's Humanity's Last Exam score, which rises from 7.17 to 13.90 in its table.^[1]

On efficiency, Liquid AI measured the model with llama.cpp at Q4_0 quantization and 32K tokens of context and said that, on all three test devices, LFM3-1.7B prefills at least 2.5 times faster and decodes at least 1.8 times faster than every compared model while using the least memory.^[3] It reported that a full 128K-token context needs 1.6 GB of KV cache, against 4.3 GB for Qwen3.5-4B and 5.6 GB for MiniCPM5-2B, and that on a single NVIDIA B300 with SGLang in BF16 the model can produce 3.6 billion tokens per day at 1,024 concurrent requests.^[1] As a throughput demonstration, the company ran more than 300,000 LFM3-1.7B agents in parallel to investigate the Enron email archive.^[1]

The launch post quoted Ivan Petrukha of MacPaw, who said that in a test of 65 everyday agent tasks in real macOS environments, LFM3-1.7B reached 64% task success against 48% for Qwen3.5-2B, used 58% fewer tokens per task, and got stuck in under 1% of attempts against 15%.^[1] Liquid AI's thread also claimed that LFM3-1.7B is "#1 on the Pareto frontier of device intelligence" on Pipette, Liquid AI's on-device model benchmarking site.^[2][13]

LFM3-VL-2B (vision)

LFM3-VL-2B accepts text, images, and video. Liquid AI describes it as the first LFM vision model to take video as input, and says it reasons step by step before answering and can localize objects and reason about their spatial relationships, which it presents as a foundation for robotics applications.^[1][5] The model card says it is designed for questions about video, reading documents, charts, and app screens, finding objects, and spatial reasoning for robots, and that it is not recommended for coding, multi-turn conversations, or complex reasoning.^[5]

Liquid AI evaluated it on 62 vision benchmarks grouped into ten categories (each category score is the unweighted mean of its benchmarks). It says the model scores at least 7.4 points ahead of every model of similar size, matches Gemma 4 12B (a model more than six times larger), surpasses Gemini 3.5 Flash Lite, and comes within 2 points of GPT-5.6 Luna.^[1] Selected Liquid AI-reported category scores:^[1]

Liquid AI also reported that LFM3-VL-2B has the highest prefill throughput and lowest memory use of the compared models of its size class on all three of its test devices, and that it produces more than 12,000 tokens per second on one NVIDIA B300 at 1,024 concurrent requests.^[1][5]

LFM3-Audio-2B (audio)

LFM3-Audio-2B takes text or audio (at 16 kHz) as input and responds with text, 24 kHz speech, or interleaved text and speech.^[6] Liquid AI lists four main capabilities: streaming voice chat (with optional reasoning for text-only replies), transcription, synthesis, and translation of speech across 18 languages, audio understanding and reasoning about speech, sounds, and events, and spoken requests turned into tool calls.^[1][6] The model card's speech-language list covers Arabic, Danish, German, Greek, British and American English, Finnish, French, Hebrew, Italian, Japanese, Korean, Dutch, Polish, Portuguese, Spanish, Swedish, Turkish, and Chinese.^[6]

Selected Liquid AI-reported results:^[1][6]

The English ASR figures are average word error rates on the ESB / Open ASR benchmark suite.^[6] On that measure the new model is roughly level with LFM2.5-Audio-1.5B; Liquid AI says LFM3-Audio-2B outperforms LFM2.5-Audio in speech quality and adds audio understanding, general-purpose audio reasoning, and tool calling as new capabilities.^[1]

LFM3-Omni-2B (omni)

LFM3-Omni-2B is Liquid AI's first omni-modal model. It accepts text, images, video with or without its soundtrack, and 16 kHz audio, and answers in text, 24 kHz speech, or both.^[1][7] The model card says one model replaces a separate speech recognizer, vision-language model, LLM, and text-to-speech model, and that it is a non-reasoning model.^[7] Liquid AI describes uses such as asking spoken questions about images or scenes, understanding video together with its audio, and combining what the model sees with spoken instructions to produce tool calls.^[1]

Liquid AI-reported results on audio-visual benchmarks, against much larger omni models:^[1]

The parameter counts in this table are the totals Liquid AI listed for each model.^[1]

d1-2B and d1-Omni-2B (decision models)

Liquid AI introduced d1, its line of "decision models," earlier in October 2026 and released the open-weight d1-3B and d1-omni-600M on October 7 (the earlier models are covered on the Liquid AI page).^[10] A decision model takes an input (a "state") and a set of questions and returns typed answers with calibrated probabilities in a single forward pass, without generating tokens.^[1][8] LFM3 adds two new d1 models: d1-2B, built on LFM3-VL-2B, and d1-Omni-2B, built on LFM3-Omni-2B, which also accepts audio and spoken questions.^[1][8][9]

Liquid AI says the main change since d1-3B, apart from the new backbone, is typed answers. In addition to yes/no, single-choice, and score questions, both models can point (draw a box around a described object or find every instance with a count), read (extract text, numbers, and dates with the span they came from, or report that a field is missing), estimate (give a number with a likely range), and pick (select every applicable option with its own probability).^[1] The model cards list routing and triage, moderation and guardrails, classification, extraction from documents and forms, judging and reranking, and visual inspection as intended uses, and state that the models are not chat models.^[8][9]

Liquid AI said d1-2B replaces d1-3B with a third fewer parameters and reported these comparisons:^[1]

The d1-2B model card reports 5.7 ms per decision on an NVIDIA H100 and 3.8 ms on an AMD Instinct MI325X.^[8] For d1-Omni-2B, Liquid AI reported that accuracy drops only modestly when questions are spoken instead of written (for example 83.4 written and 78.2 spoken on DecisionBench), along with 84.5% on the ESC-50 sound classification set and 92.5% language identification across 43 languages on FLEURS.^[1] The d1-Omni-2B card notes that long spoken passages lose much of their accuracy, especially outside Latin-script languages, and recommends d1-2B when no audio is involved.^[9]

Fine-tuning and specialization examples

Liquid AI says every LFM3 model can be fine-tuned on a single GPU, and it published notebooks using the TRL library for supervised fine-tuning, DPO, GRPO, and continued pre-training.^[1][3] The launch post included two demonstrations based on LFM3-VL-2B:^[1]

• Industrial inspection. Liquid AI fine-tuned LFM3-VL-2B to find and classify defects in drone photos of wind turbine blades, training on about 3,000 public photos from the WTBs2025 dataset with SFT and GRPO for a few hours on one A100. On about 1,400 held-out photos, it said both the base model and GPT-6 Luna identified the defect types in fewer than 10% of images, while the fine-tuned model reached 67% classification accuracy and raised detection F1 from 0.003 to 0.11, against 0.01 for GPT-6 Luna.
• Robotics. The company explored three uses. For tool orchestration, a rank-32 LoRA adapter trained on about 40,000 samples (about six GPU-hours on one AMD Instinct MI325X) let the model operate a simulated robot arm in NVIDIA Isaac Sim through pointing, depth, and motion-planning tools. For vision-and-language navigation, a version trained on 2.8 million samples reported 53% success on R2R and 43% on RxR in unseen validation scenes and updated waypoints at more than 2.5 Hz on a Jetson AGX Orin. For manipulation, Liquid AI turned the model into a vision-language-action model trained on data from 15 robot embodiments, reporting 96.2% average success across the four LIBERO suites and 45.7% on RoboCasa-GR1 tabletop tasks, with action chunks generated in under 50 ms on a Jetson AGX Orin.

Launch partners

Liquid AI named Ambarella, AMD, Arm, Intel (via OpenVINO), NVIDIA, OpenRouter, Qualcomm (via Qualcomm AI Hub and GenieX), and Vercel as day-zero launch partners, and thanked Mercedes-Benz and Shopify for pre-release testing and evaluation.^[1] Its list of supporting tools and communities included Hugging Face, llama.cpp, LM Studio, MLX, Ollama, RadixArk on SGLang, Unsloth, vLLM, and ZML, among others.^[1] The launch post carried statements from AMD, Ambarella, Qualcomm, MacPaw, and Mercedes-Benz; Ambarella's Malhar Palkar said the company had begun enabling LFM3 on its Cooper Model Garden platform, where LFM2.5-VL models already ran.^[1]

License

The LFM3 weights are released under the LFM Open License v1.0.^[3][11] The license grants commercial-use rights only to entities whose annual revenue is below a "Threshold" defined as $10 million; commercial use by a legal entity above that threshold is not licensed under the agreement. The threshold does not apply to qualified non-profit organizations using the models for non-commercial or research purposes.^[11]

Reception

Early coverage largely restated Liquid AI's own figures. Tech Times described the family as "six open-weight models" sharing "a single shared backbone of just 1.7 billion parameters" and highlighted the WebArena-Verified jump from 3.02 for LFM2.5-2.6B to 38.42 for LFM3-1.7B, while noting that all of the benchmark figures were self-reported by Liquid AI and had not been independently reproduced at the time of writing.^[12]

See also

• Liquid AI, including the LFM2 and LFM2.5 generations and the original d1 models
• Small language model
• Edge AI
• Sliding window attention

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53Liquid AI. "Introducing LFM3: Every Modality for Specialized Intelligence." Liquid AI Blog, October 27, 2026. liquid.ai/...lfm3
2. ^1 ^2 ^3 ^4 ^5 ^6Liquid AI (@liquidai). "Introducing LFM3: our new generation of Liquid Foundation Models..." X, October 27, 2026. x.com/...2115128329003303107
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Liquid AI. "LiquidAI/LFM3-1.7B" model card. Hugging Face, accessed October 29, 2026. huggingface.co/...LFM3-1.7B
4. ^1 ^2 ^3 ^4 ^5 ^6Liquid AI. "LiquidAI/LFM3-1.7B-Base" model card. Hugging Face, accessed October 29, 2026. huggingface.co/...LFM3-1.7B-Base
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Liquid AI. "LiquidAI/LFM3-VL-2B" model card. Hugging Face, accessed October 29, 2026. huggingface.co/...LFM3-VL-2B
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Liquid AI. "LiquidAI/LFM3-Audio-2B" model card. Hugging Face, accessed October 29, 2026. huggingface.co/...LFM3-Audio-2B
7. ^1 ^2 ^3 ^4 ^5Liquid AI. "LiquidAI/LFM3-Omni-2B" model card. Hugging Face, accessed October 29, 2026. huggingface.co/...LFM3-Omni-2B
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7Liquid AI. "LiquidAI/d1-2B" model card. Hugging Face, accessed October 29, 2026. huggingface.co/...d1-2B
9. ^1 ^2 ^3 ^4 ^5Liquid AI. "LiquidAI/d1-Omni-2B" model card. Hugging Face, accessed October 29, 2026. huggingface.co/...d1-Omni-2B
10. ^Liquid AI. "Open d1: Edge decision models for text, vision, and audio." Liquid AI Blog, October 7, 2026. liquid.ai/...d1-open
11. ^1 ^2Liquid AI. "LFM Open License v1.0." Hugging Face (LFM3-1.7B repository), accessed October 29, 2026. huggingface.co/...LICENSE
12. ^Rutherford, Mark. "Liquid AI Releases LFM3: Six Open-Weight Models Fuse Text, Vision, Audio on One 2B Backbone." Tech Times, October 28, 2026. techtimes.com/...text-vision-audio-one-2b-backbone
13. ^Liquid AI. "Pipette: On-Device Foundation Model Benchmarks." Accessed October 29, 2026. pipette.liquid.ai