Introducing LFM3: Every Modality for Specialized Intelligence

Liquid AI released LFM3, a new generation of Liquid Foundation Models built on a shared multimodal backbone covering text, vision, audio, omni and decision models, with the first release spanning models in the 2B parameter range. LFM3-1.7B-Base was pre-trained on 22T tokens and uses a tri-hybrid architecture of 24 layers — 13 gated short convolutions, 5 sliding-window attention layers with a 256-token window, and 6 global attention layers — that Liquid AI says makes LFM3-1.7B about 25% faster than LFM2-1.2B at 8K tokens of context per parameter. The open-weight base and post-trained models are available on Hugging Face, the Liquid AI Playground and the Liquid AI API, and Liquid AI expanded its hardware-in-the-loop architecture search from CPUs in one phone and one laptop for LFM2 to 14 device and backend combinations for LFM3.

News https://www.liquid.ai/news Models https://www.liquid.ai/news/models Introducing LFM3: Every Modality for Specialized Intelligence Today, we’re excited to announce LFM3 , our new generation of Liquid Foundation Models LFMs built for specialized intelligence. The LFM3 model family is multimodal by nature and brings text, vision, audio, omni, and decision models onto one new shared backbone. The first release of the LFM3 family covers models in the 2B parameter range. Built on a new hybrid architecture and 22T tokens of pre-training, LFM3 is far more expressive than LFM2 https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models while maintaining its efficiency. The extra capacity in a small, fast model makes LFM3 the ideal starting point in its class for specialized intelligence: this is because it can be fine-tuned locally, it can run on any device, and it is capable enough to become a reliable expert on a specialized task. The base and post-trained models are open-weight and available today on Hugging Face https://huggingface.co/collections/LiquidAI/lfm3 , our Playground https://playground.liquid.ai/ , and the Liquid AI API https://console.liquid.ai/dashboard . Check out our docs https://docs.liquid.ai/ on how to run and fine-tune them locally. Download weights https://huggingface.co/collections/LiquidAI/lfm3 Try in Liquid API https://console.liquid.ai/dashboard Try in Playground https://playground.liquid.ai/ Benchmark in Pipette https://pipette.liquid.ai/ New architecture for more expressivity LFM2 showed that a hybrid of short convolutions and attention could run fast on-device without sacrificing quality. Its design came from our automated hardware-in-the-loop architecture search https://arxiv.org/abs/2511.23404 . For LFM3, we used this search to redesign that architecture for agentic use cases, which need more capability per parameter and fast inference over long contexts. We also expanded 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 , and Arm and x86 cloud CPUs, plus GPUs and the Qualcomm Hexagon NPU. Gated short convolutions still make up over half the layers. To add capability without slowing down long contexts, we tried swapping some of them for other sequence blocks, and sliding-window attention SWA gave the best quality–speed tradeoff. We kept the other changes described below only where they improved quality without a meaningful speed cost. The payoff shows at long context. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices, and with 8K tokens in context it is about 25% faster. Tri-hybrid layer mix. LFM3 combines three operators: gated short convolutions, sliding-window attention SWA , and global attention. Of its 24 layers, 13 are convolutions, 5 are SWA with a 256-token window, and 6 are global attention. The convolutions also use a kernel size of 7, up from 3 in LFM2. Head-wise gated attention. A sigmoid gate scales each head’s output, so a head can switch off when it has nothing to add. This can help suppress activation outliers https://arxiv.org/abs/2306.12929 that hurt quantization and improve training stability https://arxiv.org/abs/2505.06708 . Partial RoPE. Global attention layers rotate only the highest-frequency half of each head and leave the low-frequency half unrotated, so those dimensions match on content regardless of distance https://arxiv.org/abs/2410.06205 . SWA layers keep full RoPE. Sandwich norm. Each block normalizes both its input and its output, which helps keep activations well-scaled. LFM3 uses the same 128K vocabulary as LFM2.5-8B-A1B https://www.liquid.ai/blog/lfm2-5-8b-a1b , built to tokenize non-Latin scripts more efficiently. Training. LFM3-1.7B-Base was pre-trained on 22T tokens. Compared with LFM2, the mix adds more scientific papers, repository-level code, a wider range of knowledge sources, and considerably more math. The final 3.5T tokens are a mid-training stage on curated, higher-quality data, with a larger share of agentic traces and diverse tool calls to prepare the model for agentic post-training. This stage also extends the context window to 256K tokens. A multimodal family for every need The first LFM3 release includes six models, all branched from the same LFM3-1.7B-Base backbone and all around 2B parameters. Four of them cover different input modalities text, vision, audio, and omni , and the last two are decision models specialized from the vision and omni models. Instead of running one large general-purpose model, you can pick the one that fits your use case, customize it, and run it on-device. For a deeper dive into the performance and inference results behind these models, refer to the respective model cards. Text LFM3-1.7B https://huggingface.co/LiquidAI/LFM3-1.7B is a text model for local agents and specialized applications with a 256K-token context window. It is post-trained to run in agentic harnesses such as Hermes Agent, Pi, and OpenClaw. It can triage your inbox, manage your calendar, research topics on the web, send you morning briefings on WhatsApp, and use tool search to find the right tool among thousands of options. Since tokens are free, dozens of agents can run continuously in the background. You can adapt it to your domains, harnesses, and tools with a short fine-tuning run on a single GPU. | Benchmark | LFM3-1.7B 1.7B | LFM2.5-2.6B 2.6B | MiniCPM5 2B 2.5B | Gemma 4 E4B 8B | Qwen3.5 4B 4.7B | Qwen3.5 9B 9.7B | | Agents | | | | | | | | LiveClaw | 45.60 | 14.70 | 30.37 | 20.36 | 19.94 | 35.77 | | Claw-Eval General EN , average | 68.64 | 62.63 | 60.74 | 56.14 | 61.96 | 66.53 | | Claw-Eval General EN , pass^3 | 51.40 | 37.52 | 34.58 | 22.43 | 42.99 | 51.87 | | BrowseComp-Plus OpenClaw subset | 37.00 | 26.33 | 16.83 | 21.00 | 24.67 | 38.00 | | WebArena-Verified | 38.42 | 3.02 | 9.11 | 17.06 | 21.92 | 28.94 | | EnterpriseOps-Gym | 28.20 | 7.94 | 13.41 | 9.86 | 25.96 | 28.97 | | AutomationBench public | 41.29 | 0.94 | 4.71 | 1.92 | 8.67 | 15.10 | | Tool use | | | | | | | | BFCL V4 | 71.39 | 55.89 | 62.65 | 48.25 | 55.12 | 62.83 | | ToolSandbox | 80.36 | 78.04 | 68.16 | 69.27 | 75.40 | 76.41 | | MCP-Atlas | 32.57 | 16.45 | 26.38 | 6.19 | 41.37 | 42.67 | | τ³-bench | 68.01 | 56.77 | 69.44 | 33.53 | 65.98 | 67.11 | | Agentic coding | | | | | | | | SWE-bench Verified OpenCode | 43.40 | 18.00 | 34.80 | 0.50 | 41.20 | 51.20 | | SWE-bench Multilingual OpenCode | 23.00 | 8.00 | 15.67 | 0.67 | 23.33 | 35.00 | | Terminal-Bench 2.1 OpenCode | 12.92 | 0.00 | 5.62 | 0.00 | 13.48 | 17.98 | | Coding | | | | | | | | LiveCodeBench v6 | 73.46 | 60.57 | 70.52 | 71.00 | 61.99 | 69.86 | | LiveCodeBench Pro 25Q2 | 31.74 | 19.16 | 29.34 | 29.94 | 24.25 | 34.43 | | SciCode | 28.47 | 16.32 | 27.78 | 25.00 | 23.02 | 31.60 | | Instruction following | | | | | | | | IFEval | 96.66 | 95.39 | 87.83 | 90.02 | 92.52 | 92.08 | | IFBench | 73.03 | 60.48 | 62.08 | 44.56 | 38.16 | 48.84 | | IFStruct | 93.72 | 82.18 | 72.65 | 78.45 | 87.05 | 89.20 | | Multi-IF | 89.10 | 84.80 | 73.19 | 81.03 | 80.32 | 84.44 | | Knowledge | | | | | | | | GPQA Diamond | 63.37 | 57.78 | 65.15 | 57.88 | 75.76 | 80.40 | | Humanity's Last Exam no tools | 7.17 | 5.88 | 8.20 | 4.77 | 12.84 | 16.41 | | Humanity's Last Exam with tools | 13.90 | 9.60 | 11.50 | 6.00 | 11.60 | 17.50 | | AA-Omniscience public | 0.50 | -33.17 | -5.00 | -15.33 | -57.17 | -28.00 | | AA-Omniscience accuracy public | 7.07 | 8.75 | 9.13 | 9.83 | 15.67 | 13.17 | | AA-Omniscience non-hallucination public | 92.93 | 53.69 | 84.44 | 72.09 | 13.64 | 52.59 | | Long context | | | | | | | | AA-LCR | 52.83 | 6.00 | 49.34 | 39.33 | 52.67 | 62.00 | | RULER 128K | 65.56 | 1.15 | 37.96 | 18.82 | 84.63 | 87.11 | | Average | 49.37 | 31.55 | 39.19 | 31.74 | 39.27 | 47.24 | On agentic tasks, LFM3-1.7B competes with models several times its size. It leads on instruction following and function calling on BFCL and ToolSandbox, and beats Qwen3.5-9B on long tasks in browsers and business apps WebArena-Verified, AutomationBench . Larger models still win on agentic coding and knowledge, where size matters most, but search helps close the knowledge gap, nearly doubling its score on Humanity's Last Exam. To achieve these results, the model went through four stages of post-training: - Supervised fine-tuning SFT : 1.75T tokens across various tasks, including chat, reasoning, tool calling, and agentic traces. - Mixed reinforcement learning: many domains trained at once, from math and code to SQL and structured outputs. A length penalty in the reward keeps answers concise. - Multi-domain on-policy distillation MOPD : six specialist RL teachers code, math, tool use, long context, agentic work, and enterprise workflows are distilled back into a single model. - Agentic RL: tasks from web browsing to office work are run inside real agent harnesses, preparing the model for deployment. On-device, we measure metrics with llama.cpp at Q4 0 and a 32K-token context. LFM3-1.7B is the fastest model on all three devices, decoding 1.8 to 2.8 times faster than the next one, and it uses the least memory. Long contexts stay cheap: 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. On a single NVIDIA B300 with SGLang in BF16, LFM3-1.7B can produce 3.6 billion tokens per day at 1,024 concurrent requests, more than any other model in this comparison. To show what this throughput makes possible, we ran more than 300,000 LFM3-1.7B agents in parallel to investigate the Enron email archive. LFM3-1.7B can be served at low latency under random traffic, enabling real-time use cases like online fraud detection when run on Liquid RIPL https://www.liquid.ai/blog/low-latency . P95 request latency remains stable even under heavy load, outperforming competing models served on vLLM. Vision LFM3-VL-2B https://huggingface.co/LiquidAI/LFM3-VL-2B can watch, reason, and understand the physical world: - Video understanding: for the first time, an LFM vision model takes video as input. Ask questions about a recorded lecture, a walkthrough filmed on your phone, or a clip from a camera feed. - Thinking: the model reasons step by step before answering. This helps it better understand user inputs and solve harder visual questions such as math, charts, and multi-step problems. - Spatial and embodied reasoning: the model’s ability to localize objects and reason about their spatial relationships provides a strong foundation for downstream robotics applications. LFM3-VL-2B is a leap in performance over the LFM2.5-VL series. Across 62 vision benchmarks, LFM3-VL-2B scores significantly ahead of all models of similar size by at least 7.4 points, and despite its size, it matches the performance of Gemma 4 12B, a model more than 6 times larger, surpasses Gemini 3.5 Flash Lite and is within 2 points of GPT-5.6 Luna. For more det