Liquid AI's LFM2.5-350M Beats Larger Models Running Under 1GB
Liquid AI's LFM2.5-350M packs 28T-token training and reinforcement learning into a sub-1GB model built for edge agentic pipelines
Takeaways
- LFM2.5-350M released: Liquid AI's 350M parameter model trained on 28T tokens with scaled RL, available now on Hugging Face.
- Non-Transformer architecture: Hybrid of 10 LIV convolution blocks + 6 GQA attention blocks, drastically reducing KV cache and memory usage.
- Edge-first performance: 313 tok/s on AMD CPU, 188 tok/s on Snapdragon Gen4, under 1GB memory; 40.4K tok/s on a single H100.
- Beats larger models on agentic tasks: Outperforms Gemma 3 1B and Qwen3.5-0.8B on IFEval, IFBench, BFCLv3, and CaseReportBench.
- Not for math or code: Explicitly not recommended for knowledge-intensive tasks, programming, or creative writing.
- Open-weight, multi-format: Ships as GGUF, MLX, ONNX, OpenVINO, and vLLM with day-one support across AMD, Qualcomm, Intel, and Apple Silicon.
The dominant assumption in AI has been simple: more parameters, more intelligence. Liquid AI is pushing back hard. With the release of LFM2.5-350M, they've taken a 350-million parameter model, trained it on 28 trillion tokens with large-scale reinforcement learning, and produced something that beats models more than twice its size on several key benchmarks -- all while running under 1GB of memory.
This is not a general-purpose reasoning model. It's a precision tool: purpose-built for tool use, data extraction, and structured outputs at the edge, where compute and memory are tight and latency is non-negotiable.
Not a Transformer Under the Hood
The architecture is the first thing to understand. The core technical differentiator of LFM2.5-350M is its departure from the pure Transformer architecture -- it uses a hybrid structure built on Linear Input-Varying Systems (LIVs). LIVs are a class of sequence model that behave somewhat like advanced RNNs, but are designed to be more parallelizable and stable during training.
The hybrid backbone consists of 10 Double-Gated LIV Convolution blocks that handle the majority of sequence processing. LIVs function similarly to advanced RNNs but are designed to be more parallelizable and stable during training, maintaining a constant-state memory that reduces I/O overhead. The model also includes 6 Grouped Query Attention (GQA) blocks, which let it retain high-precision retrieval and long-range context handling without the full memory overhead of a standard Transformer -- supporting a 32k context window while maintaining an extremely lean memory footprint.
The practical consequence: because LFM2.5-350M uses LIVs and GQA, it drastically reduces KV cache size, boosting throughput. The KV cache is the memory structure Transformers use to store past context during generation -- it grows linearly with sequence length and is a major bottleneck at the edge.