Member of Technical Staff

Insilico Medicine and Liquid AI have partnered to create LFM2-2.6B-MMAI, a lightweight scientific foundation model for drug discovery that runs entirely on private pharmaceutical infrastructure. The 2.6-billion-parameter model achieves state-of-the-art performance across multiple drug discovery tasks whilst being ten times smaller than comparable systems. The model covers property prediction, molecular optimisation, affinity prediction and chemical reasoning. It outperformed TxGemma-27B on 13 of 22 pharmacokinetics and toxicology tasks, achieved 98.8% success rates on multi-parameter optimisation benchmarks, and produced better correlation scores than GPT-5.1, Claude Opus 4.5 and Grok-4.1 on Insilico's internal benchmark featuring 2.5 million experimental measurements across 689 protein targets. The collaboration addresses pharmaceutical companies' need to use advanced AI capabilities without sending proprietary data to external cloud services.

Liquid AI releases LFM2.5-1.2B-Thinking: a 1.2B parameter reasoning model that fits under 1 GB on-device. Liquid AI has released LFM2.5-1.2B-Thinking, a 1.2 billion parameter reasoning model that runs fully on device and fits in about 900 MB on a modern phone. What needed a data center 2 years ago can now run offline on consumer hardware, with a focus on structured reasoning traces, tool use, and math, rather than general chat. Position in the LFM2.5 family and core specs. LFM2.5-1.2B-Thinking is part of the LFM2.5 family of Liquid Foundation Models, which extends the earlier LFM2 architecture with more pre-training and multi stage reinforcement learning for edge deployment. The model is text only and general purpose with the following configuration: * 1.17B parameters, reported as a 1.2B class model * 16 layers, with 10 double gated LIV convolution blocks and 6 GQA blocks * Training budget of 28T tokens * Context length of 32,768 tokens * Vocabulary size of 65,536 * 8 languages, English, Arabic, Chinese, French, German, Japanese, Korean, Spanish Reasoning first behavior and thinking traces. The 'Thinking' variant is trained specifically for reasoning. At inference time it produces internal thinking traces before the final answer. These traces are chains of intermediate steps that the model uses to plan tool calls, verify partial results, and work through multi step instructions. Liquid AI team recommends this model for agentic tasks, data extraction pipelines, and retrieval augmented generation flows where you want explicit reasoning and verifiable intermediate steps. A practical way to think about it, you use LFM2.5-1.2B-Thinking as the planning brain inside agents and tools, and use other models when you need broad world knowledge or code heavy workflows. Benchmarks versus other 1B class models. Liquid AI team evaluates LFM2.5-1.2B-Thinking against models around 1B parameters on a suite of reasoning and instruction benchmarks. Compared to LFM2.5-1.2B-Instruct, three metrics improve strongly, math reasoning rises from about 63 to 88 on MATH 500, instruction following rises from about 61 to 69 on Multi IF, and tool use rises from about 49 to 57 on BFCLv3. LFM2.5-1.2B-Thinking competes with Qwen3-1.7B in thinking mode on most reasoning benchmarks while using around 40 percent fewer parameters and fewer output tokens on average. It also outperforms other 1B class baselines such as Granite-4.0-H-1B, Granite-4.0-1B, Gemma-3-1B-IT, and Llama-3.2-1B Instruct on many of these tasks. Training recipe and doom looping mitigation. Reasoning models often suffer from doom looping, where the model repeats fragments of its chain of thought instead of finishing the answer. LFM2.5-1.2B-Thinking uses a multi stage training pipeline to reduce this. The process starts with mid training that includes reasoning traces so the model learns a 'reason first then answer' pattern. Then supervised fine tuning on synthetic chains improves chain of thought generation. After that, preference alignment and RLVR are applied. In preference alignment, the research team generates 5 temperature sampled candidates and 1 greedy candidate per prompt and uses an LLM judge to pick preferred and rejected outputs, while also labeling looping outputs explicitly. During RLVR they add an n gram repetition penalty early in training. This reduces the doom loop rate from 15.74 percent at mid training to 0.36 percent after RLVR on a set of representative prompts. The result is a small reasoning model that can produce thinking traces without getting stuck in long repetitive outputs, which is important for interactive agents and on device UX. Inference performance and hardware footprint. A key design target is fast inference with a small memory footprint on CPUs and NPUs. LFM2.5-1.2B-Thinking can decode at about 239 tokens per second on an AMD CPU and about 82 tokens per second on a mobile NPU, while running under 1 GB of memory, with broad day one support for llama.cpp, MLX, and vLLM. The detailed hardware table uses 1K prefill and 100 decode tokens and gives the following examples for LFM2.5-1.2B-Thinking These numbers show that the model fits comfortably under 1 GB on phones and embedded devices while sustaining useful throughputs even at long contexts. Key takeaways. * LFM2.5-1.2B-Thinking is a 1.17B parameter reasoning model with 32,768 context length and runs under 1 GB on phones and laptops. * The model is optimized for explicit thinking traces, agentic workflows, data extraction, and RAG. * It reaches strong scores for a 1B class model, for example 87.96 on MATH 500, 85.60 on GSM8K, and competitive performance with Qwen3 1.7B in thinking mode with fewer parameters. * The training pipeline uses midtraining with reasoning traces, supervised fine tuning, preference alignment with 5 sampled along with 1 greedy candidate, and RLVR with n gram penalties, which reduces doom loops from 15.74 percent to 0.36 percent. * The model runs efficiently on AMD and Qualcomm NPUs and CPUs with runtimes like llama.cpp, FastFlowLM, and NexaML, is available in GGUF, ONNX, and MLX formats, and can be loaded easily from Hugging Face for on device deployment. Hosting providers/deployment. You can access or host the model through the following providers and platforms: Cloud & API providers. Model repositories (self-hosting). If you want to run the model locally or on your own infrastructure, the weights are available in various formats:

Liquid AI LFM2.5-1.2B-Thinking: compact power. Exploring Liquid AI's newest 1.2B reasoning model optimized for agentic tasks, RAG, and high-speed edge inference with LIV convolution architecture. Liquid AI LFM 2.5 Edge AI On-device AI Compact Models Agentic AI RAG Machine Learning Liquid AI has recently unveiled its latest breakthrough in compact language models: the LFM2.5-1.2B-Thinking. As part of the LFM 2.5 family, this model represents a significant step forward in bringing sophisticated reasoning capabilities to edge devices. Unlike traditional Transformers that struggle with memory and compute constraints on smaller hardware, Liquid's models leverage a unique architecture designed for efficiency and speed. The LFM2.5-1.2B-Thinking is specifically tuned for precision and logic, making it a powerful tool for developers looking to build local agents or fast RAG pipelines without the need for massive GPU clusters. Key features and architecture. The "1.2B" in its name refers to its 1.17 billion parameters, yet its performance punches well above its weight class. Here are the core specifications: * Liquid Architecture: The model consists of 16 layers, utilizing 10 double-gated LIV (Liquid Interleaved Variable) convolution blocks paired with 6 GQA (Grouped Query Attention) blocks. * Large Context Window: Despite its size, it supports a 32,768 token context length, which is essential for complex RAG tasks and long-form data extraction. * Massive Training Data: Liquid AI trained this model on a staggering 28 trillion tokens, ensuring a high degree of "world knowledge" and linguistic fluency for its size. * Multilingual Support: Out of the box, it supports English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish. Performance & optimization. One of the most impressive aspects of the LFM2.5-1.2B-Thinking is its inference speed. Liquid AI has partnered with industry leaders like AMD and Qualcomm to optimize the LFM family for NPUs (Neural Processing Units). * CPU & NPU Efficiency: The model offers extremely fast inference on standard CPUs with low memory footprints. * Edge Performance: On AMD Ryzen NPUs using FastFlowLM, the model can sustain ~52 tokens per second at 16K context and ~46 tokens per second even at its full 32K context. * Compact Thinking: Compared to models like Qwen3-1.7B, the LFM2.5-1.2B-Thinking achieves comparable or better results while requiring fewer output tokens to reach the same conclusion. Use cases: agents and RAG. Liquid AI recommends this model for specific scenarios where speed and reliability are paramount: * Agentic Tasks: Its ability to handle "thinking" steps and function calling makes it ideal for autonomous agents that need to run locally. * Data Extraction: Its reasoning capabilities allow it to parse complex documents and extract structured information with high accuracy. * Local RAG: With a 32K context window and fast inference, it's perfect for searching through local knowledge bases and summarizing information on-the-fly. [!IMPORTANT] While powerful, Liquid AI notes that this model is not recommended for knowledge-intensive "Jeopardy-style" tasks or deep programming work, where larger-scale models still hold the edge. The release of LFM2.5-1.2B-Thinking signals a shift in the AI industry toward specialized, compact models. By focusing on architecture-level efficiency rather than just raw size, Liquid AI is making "thinking" AI accessible for vehicles, mobile devices, and IoT hardware. As on-device AI continues to evolve, models like this will be the backbone of the next generation of privacy-focused, always-on digital assistants.