Liquid AI has released the LFM2.5 model family, a comprehensive suite of open-weight models designed to run entirely on edge devices without cloud dependencies. The announcement spans two major releases: the LFM2.5-1.2B family introduced in January 2026 and the larger LFM2.5-2.6B agentic model released in August 2026. Both model lines share a hybrid architecture that combines multiplicative gates with short convolutions, enabling faster inference and lower memory footprints than comparable transformer-based models at similar parameter counts.
The company positions these releases as a step toward "AI that runs anywhere," emphasizing private, fast, and always-on intelligence for smartphones, vehicles, IoT devices, and embedded systems. All models are available under open-weight licenses on Hugging Face and through Liquid's LEAP platform, with day-one support across llama.cpp, MLX, vLLM, ONNX, and SGLang inference frameworks. Partnerships with AMD and Nexa AI bring optimized NPU deployment for the 1.2B variants, while the 2.6B model demonstrates competitive agentic performance against models nearly four times its size.
What's New / Specs
The LFM2.5-1.2B family comprises five specialized variants built on a shared pretrained backbone. The Base checkpoint serves as the foundation for all downstream models, having been pretrained on 28 trillion tokens — a significant increase from the 10 trillion tokens used for the original LFM2 series. The Instruct variant adds supervised fine-tuning, preference alignment, and large-scale multi-stage reinforcement learning to deliver strong instruction following and tool use capabilities out of the box.
- LFM2.5-1.2B-Base: Pretrained checkpoint (28T tokens) recommended for heavy fine-tuning, domain-specific assistants, and proprietary data training.
- LFM2.5-1.2B-Instruct: General-purpose instruction-tuned model with multi-stage RL post-training for instruction following and tool use.
- LFM2.5-1.2B-JP: Japanese-optimized chat model achieving state-of-the-art results on JMMLU (50.7), M-IFEval (58.1), and GSM8K-JP (56.0) at the 1B scale.
- LFM2.5-VL-1.6B: Vision-language model with improved multi-image comprehension and multilingual vision understanding across Arabic, Chinese, French, German, Japanese, Korean, and Spanish.
- LFM2.5-Audio-1.5B: Native audio-language model with an LFM-based detokenizer that is 8x faster than the LFM2 Mimi detokenizer on mobile CPU, supporting INT4 quantization-aware training with minimal quality loss.
The LFM2.5-2.6B model, released separately in August 2026, targets agentic workloads with a four-stage post-training pipeline: Supervised Fine-Tuning (SFT), Teacher Specialization, Multi-Domain On-Policy Distillation (MOPD), and Agentic Reinforcement Learning. It was pretrained on approximately 34 trillion tokens with a vocabulary expanded to 128K tokens (doubled from 64K) through in-place tokenizer extension. A dedicated 128K context-extension phase during mid-training enables handling the long inputs typical of agentic workflows. The model ships with both Base and post-trained variants on Hugging Face.
Benchmark results published by Liquid AI show the 1.2B-Instruct model leading its size class across knowledge (GPQA 38.89, MMLU-Pro 44.35), instruction following (IFEval 86.23, IFBench 47.33, Multi-IF 60.98), math (AIME25 14.00), and tool use (BFCLv3 49.12) benchmarks, outperforming Llama 3.2 1B Instruct, Gemma 3 1B IT, and Granite 4.0 1B models. The 2.6B model leads on instruction-following benchmarks (IFBench 59.17, Multi-IF 80.07, IFStruct 85.49) and most tool-use benchmarks against larger competitors including Gemma 4 E2B/E4B (5.1B/8B) and Qwen 3.5 4B/9B, trailing only Qwen 3.5 9B on BFCLv4 and coding benchmarks.
Why It Matters
The LFM2.5 releases address three critical constraints that have limited on-device AI adoption: model quality at small parameter counts, inference efficiency on heterogeneous hardware, and the operational complexity of deploying across diverse edge environments. By extending pretraining to 28-34 trillion tokens and investing heavily in multi-stage reinforcement learning post-training, Liquid AI demonstrates that sub-3B models can achieve capabilities previously associated with 7-10B parameter models, particularly for instruction following, tool use, and agentic workflows.
Deployment flexibility is a core design principle. The models ship with GGUF checkpoints for llama.cpp (CPU inference), MLX-optimized weights for Apple Silicon unified memory, vLLM and SGLang support for GPU-accelerated serving, and ONNX exports for cross-platform accelerator support spanning AMD, Qualcomm, and Nvidia hardware. The AMD and Nexa AI partnerships specifically target NPU acceleration: benchmarks show the 1.2B-Instruct model achieving 2,591 prefill and 63 decode tokens/second on a Snapdragon X Elite NPU via NexaML, and 4,391 prefill / 82 decode tokens/second on a Snapdragon Gen4 NPU (ROG Phone 9 Pro). On mobile CPUs, llama.cpp Q4_0 quantization delivers 335 prefill / 70 decode tokens/second on a Snapdragon Gen4 (Galaxy S25 Ultra) with a 719MB memory footprint — compared to 1,306MB for Qwen3-1.7B at 181/40 tokens/second.
The audio model's 8x faster detokenizer and INT4 quantization-aware training address a persistent bottleneck in on-device speech applications: end-to-end latency. By processing audio natively rather than through pipelined ASR-LLM-TTS stages, LFM2.5-Audio-1.5B eliminates information loss between components while maintaining quality (STOI 0.89, UTMOS 3.53 at INT4 vs. 0.89/3.65 for LFM2 Mimi at FP32). The vision model's multilingual gains across seven languages expand the practical utility of on-edge multimodal AI for global deployments where cloud connectivity is unreliable or privacy regulations prohibit data egress.
Our Take
Liquid AI's strategy of scaling pretraining compute and post-training sophistication at fixed parameter budgets represents a pragmatic path for edge AI. The benchmark gains are real and measurable, but they come with caveats: the 2.6B model still trails larger models on coding-heavy workloads (LiveCodeBench v6: 59.41 vs. Qwen 3.5 9B's 69.86), and the company acknowledges that "for more complex agentic tasks or coding-heavy workloads, larger models may still be a better fit." This transparency is welcome in a space often dominated by overstated parity claims.
The open-weight release under permissive terms, combined with day-one framework support, lowers the barrier for developers to experiment with local agents. However, the practical developer experience will depend on ecosystem maturation: documentation quality, fine-tuning tooling for the hybrid architecture, and community-contributed quantizations for niche hardware. The LEAP platform's promise of "deploying models to iOS and Android as easily as calling a cloud API" remains to be validated at scale. For now, the LFM2.5 family offers the most complete open-weight package for on-device agentic AI we've seen, but its long-term impact hinges on whether the architecture's efficiency advantages translate into sustained developer adoption beyond the initial benchmark cycle.
FAQ
What are the key differences between LFM2.5-1.2B and LFM2.5-2.6B?
The 1.2B family focuses on breadth with five specialized variants (Base, Instruct, Japanese, Vision-Language, Audio-Language) pretrained on 28T tokens. The 2.6B model is a single agentic-focused variant pretrained on 34T tokens with a 128K vocabulary and 128K context window, using a four-stage post-training pipeline (SFT, Teacher Specialization, MOPD, Agentic RL) optimized for planning, tool calling, and multi-step tasks. The 2.6B model targets higher-complexity agentic workloads while the 1.2B family covers diverse modalities at lower resource requirements.
Can these models run on a smartphone without an internet connection?
Yes. Both model families are designed for fully local inference. Benchmarks show LFM2.5-1.2B-Instruct running at 30+ tokens/second on mobile CPUs (Snapdragon Gen4, Galaxy S25 Ultra) with a ~719MB memory footprint using llama.cpp Q4_0 quantization. The 2.6B model maintains ~30 tokens/second on phones and 220 tokens/second on an M5 Max. No cloud API calls are required for inference once the model weights are downloaded.
What inference frameworks are supported at launch?
Day-one support includes LEAP (Liquid's Edge AI Platform for iOS/Android), llama.cpp (GGUF checkpoints for CPU), MLX (Apple Silicon), vLLM (GPU serving), ONNX (cross-platform accelerators), and SGLang (GPU serving, added for 2.6B). All frameworks support CPU and GPU acceleration across Apple, AMD, Qualcomm, and Nvidia hardware. NPU optimization is available via AMD and Nexa AI partnerships for the 1.2B variants.
How does the audio model achieve 8x faster detokenization?
LFM2.5-Audio-1.5B replaces the LFM2 Mimi detokenizer with a custom LFM-based architecture that efficiently converts discrete language model tokens into high-fidelity audio waveforms. This new detokenizer was quantization-aware trained at INT4 precision, enabling deployment at low precision with minimal quality loss (STOI 0.89, UTMOS 3.53 at INT4 vs. 0.89/3.65 for LFM2 Mimi at FP32). The native audio processing eliminates the latency of pipelined ASR-LLM-TTS approaches.
Are there any licensing restrictions on commercial use?
The models are described as "open-weight — download, fine-tune, and deploy without restrictions" in Liquid AI's announcements. Both the 1.2B and 2.6B families are available on Hugging Face under open-weight terms. However, developers should verify the specific license files in each Hugging Face repository for any usage conditions, as the company's blog posts do not reproduce the full license text.
