LFM2.5-1.2B-Thinking
LFM2.5-1.2B-Thinking is a 1.2 billion parameter open weight reasoning model by Liquid AI that runs entirely on device under 1GB memory. It delivers advanced mathematics, tool use, and instruction following capabilities for edge AI deployment.
About LFM2.5-1.2B-Thinking
LFM2.5-1.2B-Thinking is a compact reasoning model designed to bring data center level AI capabilities to mobile and edge devices. With just 1.2 billion parameters, it fits within 900MB of memory on smartphones while generating explicit thinking traces to solve complex problems systematically. The model represents a significant advancement in on device AI, enabling sophisticated reasoning without internet connectivity or cloud dependencies.
Developed by Liquid AI, this model employs a unique curriculum reinforcement learning approach that trains domain specific capabilities in parallel rather than simultaneously. This methodology prevents capability interference while optimizing for mathematics, tool use, and advanced reasoning. The model achieves superior performance compared to larger alternatives, requiring 40% fewer parameters while delivering faster inference and more concise outputs.
The ecosystem support is extensive, with day zero compatibility across Qualcomm Snapdragon NPUs, AMD Ryzen AI processors, Apple Silicon, and NVIDIA GPUs. Native integration with popular frameworks like llama.cpp, MLX, vLLM, and Ollama ensures developers can deploy immediately across vehicles, smartphones, laptops, IoT devices, and embedded systems.
Key Features
- On Device Reasoning: Generates explicit thinking traces before producing answers, enabling systematic problem solving for mathematics and logic entirely offline.
- Ultra Low Memory: Fits within 900MB on phones and 720MB to 853MB on typical deployments, making advanced AI accessible for resource constrained environments.
- Curriculum RL Training: Uses parallel domain specific tracks with iterative model merging to optimize reasoning, math, and tool use capabilities without interference.
- Doom Loop Prevention: Reduces repetitive generation patterns from 15.74% to 0.36% through advanced preference alignment and n gram repetition penalties.
- Long Context Support: Maintains robust performance up to 32K context length with sustained decoding throughput of 52 tok/s at 16K context on NPUs.
- Broad Hardware Ecosystem: Native support for Qualcomm Hexagon, AMD XDNA, and Apple Neural Engine via partnerships with Nexa AI and FastFlowLM.
Pricing
-
Open Weight: $0 Free to download, fine tune, and deploy without restrictions from Hugging Face and Liquid AI repositories.
-
Enterprise: Contact sales Custom pricing and solutions available for enterprise deployments requiring additional support or specialized implementations.
Pricing last updated: February 26, 2026 at 11:20 AM
Use Cases
- On device AI assistants with privacy preserving reasoning capabilities
- Edge computing applications in vehicles and IoT devices
- Mathematical computation and programming assistance offline
- Agentic workflows requiring tool use and multi step planning
Pros & Cons
Pros:
- Runs entirely offline on mobile devices with sub 1GB memory
- Delivers superior performance with 40% fewer parameters than comparable models
- Generates concise outputs requiring fewer tokens than competitors
- Extensive hardware ecosystem with NPU optimization for Qualcomm and AMD
Cons:
- Not optimized for creative writing or general chat applications
- Requires specific frameworks for optimal NPU performance
Integrations
llama.cpp, MLX, vLLM, ONNX Runtime, Ollama, FastFlowLM, Cactus Engine, NexaSDK, LM Studio, TRL, Unsloth
FAQ
Explore rich media demos, screenshots, and walkthroughs for LFM2.5-1.2B-Thinking.
View MediaLast edited
February 26, 2026 at 11:20 AM by Venkatraman
Similar to LFM2.5-1.2B-Thinking
