LFM2.5-2.6B Requirements: 2.5GB RAM, 128K Context, On-Device

Liquid AI released LFM2.5-2.6B on August 4, 2026, an on-device, agent-focused model with roughly 2.69B parameters. Its official memory footprint stays under 2.5GB, and it has reportedly been run on smartphones and Raspberry Pi-class hardware. It ships under Liquid AI's own LFM1.0 license, not Apache 2.0, so commercial use terms need to be checked against the official license text. Liquid AI previously covered a Japanese-focused release in Liquid AI's two Japanese-tuned models explained, and LFM2.5-2.6B extends that lineup as a general-purpose agent model.

The FP16/Q8_0/Q4_K_M rows are standard quantization-sizing estimates derived from the 2.69B parameter count; actual memory use varies by inference engine and context length setting. Liquid AI's stated "under 2.5GB" figure applies to its own distributed build and matches the reports of it running on smartphones and Raspberry Pi-class devices.

LFM2.5-2.6B uses a 30-layer hybrid architecture combining 22 double-gated short convolution blocks with 8 GQA (Grouped Query Attention) layers. It supports a 131,072-token (128K) context window with a 128,000-token vocabulary, and was trained on roughly 34 trillion tokens. It covers 16 languages — English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, and Polish. The design goal is to run agentic workloads — tool calling, multi-step execution, web search — entirely on-device without relying on the cloud.

Liquid AI reports that LFM2.5-2.6B (77.83) outscored Qwen3.5-9B (76.44) on the ToolSandbox benchmark despite the latter being roughly four times larger. Comparisons also included Gemma-4-E2B (5.1B), Gemma-4-E4B (8B), and Qwen3.5-4B (4.7B), with Liquid AI claiming a lead across instruction-following benchmarks against all of them. Detailed per-model scores beyond what's published are not available, and no independent third-party verification has been confirmed. Teams evaluating adoption should run their own benchmarks against real workloads before relying on these figures.

- Ollama: One-command setup, good for prototyping agents and personal use
- LM Studio: GUI-based model management and chat testing, approachable for non-engineers
- llama.cpp: Runs GGUF builds directly, well suited to CPU-only or embedded deployments
- MLX: Optimized for Apple Silicon, best choice for fast inference on a Mac
- vLLM / SGLang: Chosen when server-side deployments need high throughput and heavy parallelism

Four distribution formats are available — native Transformers, GGUF (for llama.cpp), ONNX, and MLX (for Apple Silicon) — so you can pick the one that fits your runtime. Server deployments are supported through vLLM and SGLang, while local deployments run on llama.cpp, LM Studio, Ollama, and Transformers.

LFM2.5-2.6B is not released under a generic open-source license like Apache 2.0 — it ships under Liquid AI's own LFM1.0 license. Whether and how it can be used commercially, including any usage-scale restrictions or redistribution terms, is defined by that license text. Teams considering it for production should read the official license document directly and, where needed, route it through legal review before committing.

- Good fit for: agentic tasks involving tool calling, multi-step workflow execution, web-search integration, structured data extraction, and lightweight on-device RAG
- Deployment under memory constraints: the sub-2.5GB footprint makes always-on execution feasible on smartphones, edge devices, and Raspberry Pi-class hardware
- Limitations: at 2.69B parameters, it can fall behind larger models on complex multi-step reasoning or tasks requiring broad world knowledge
- Verification needed: because benchmark figures are self-reported by Liquid AI with no confirmed independent verification, evaluate against your own data before production rollout

LFM2.5-2.6B pairs a sub-2.5GB memory footprint with a 128K-token context window, positioning it as an on-device, agent-focused model. Liquid AI reports that it outscored Qwen3.5-9B on the ToolSandbox benchmark in its own comparisons, making it worth a look for lightweight agentic use cases involving tool calling and multi-step execution. If you're evaluating similar-class models, it's worth comparing against the complete Gemma 4 E4B guide and building an internal knowledge search with Qwen3.5-9B and RAG to find the best fit for your workload. Confirm the license terms and validate benchmarks against your own data before moving to production.