Liquid AI Ships LFM2.5-2.6B, an On-Device Agent Model

Introduction

On August 4, 2026, Liquid AI released LFM2.5-2.6B, a 2.6-billion-parameter language model built to run autonomous agent workflows entirely on-device — phones, laptops, and even robots — without a cloud API call. Both a base checkpoint and a post-trained, instruction-tuned checkpoint are live on Hugging Face under Liquid AI's LFM1.0 license. The release continues the company's push toward small models that plan, call tools, and complete multi-step tasks locally, following its earlier LFM2-24B-A2B hybrid-architecture model released in February 2026.

Feature Overview

LFM2.5-2.6B totals 2.69 billion parameters across 30 layers, mixing 22 double-gated short-convolution blocks with 8 grouped-query-attention (GQA) blocks, and uses a 128,000-token vocabulary with support for 16 languages. It carries a 131,072-token (128K) context window, unusually long for a model this size, which supports long agent traces and document-heavy retrieval-augmented generation without truncation.

The model was pre-trained on roughly 34 trillion tokens, then post-trained through a four-stage pipeline: supervised fine-tuning, teacher specialization, multi-domain on-policy distillation (MOPD), and agentic reinforcement learning using GRPO inside sandboxed agent environments. On Liquid AI's ToolSandbox benchmark, LFM2.5-2.6B scores 77.83, ahead of the sub-10B competitors the company tested — including Gemma-4 and Qwen3.5 — which ranged from 52.40 to 76.44 on the same test. It also leads or trades closely with similarly sized models on AIME25, LiveCodeBench, and IFBench.

Deployment support is broad from day one: the model works across llama.cpp, MLX, vLLM, SGLang, and ONNX, running on Apple Silicon, AMD, Qualcomm, and NVIDIA hardware without additional porting work.

Usability Analysis

On an Apple M5 Max, the model decodes at roughly 220 tokens per second; on an AMD Ryzen AI Max+ 395 CPU, about 113 tokens per second; and on a phone, roughly 30 tokens per second — all inside a memory footprint under 2.5GB. On a single NVIDIA H100 GPU at high concurrency, throughput scales to around 15,000 output tokens per second, meaning the same weights serve both edge devices and server-side batch inference.

Because everything runs on-device, data never leaves the device and the marginal cost per run is close to zero, which matters for privacy-sensitive agent use cases such as personal assistants and robotics, where a round-trip to a cloud API is undesirable or simply not feasible. Liquid AI is explicit about scope, though: it does not recommend LFM2.5-2.6B for agentic coding or knowledge-heavy tasks. The model is positioned for tool use, planning, data extraction/RAG, and long-context workflows — not as a general-purpose replacement for a large frontier model.

Pros & Cons

Pros

• Fully on-device agentic inference with near-zero marginal cost and strong data-privacy properties
• Broad day-one runtime and hardware support across five inference engines and four hardware vendors
• Competitive ToolSandbox benchmark result against larger sub-10B rivals
• Unusually long 128K context window for a 2.6B-parameter model
• Both base and post-trained checkpoints released openly for further fine-tuning

Cons

• The LFM1.0 license imposes usage restrictions rather than a fully permissive open-source license
• Liquid AI itself excludes agentic coding and knowledge-heavy tasks from the model's recommended use cases — a narrow specialist, not a general-purpose model
• On-phone throughput (about 30 tokens/second) trails desktop and CPU performance and may limit latency-sensitive use cases

Outlook

Small, on-device agent models are becoming a distinct product category as phone and PC makers push local AI features, and Liquid AI's four-stage distillation-plus-RL recipe is a template other small-model labs are likely to converge on. With broad runtime support already in place, LFM2.5-2.6B's near-term test is third-party adoption: how many robotics, assistant, and mobile-app developers actually swap a cloud-hosted agent loop for a fully local one, and how the model's license terms shape adoption among larger commercial deployments.

Conclusion

LFM2.5-2.6B is a well-scoped, well-benchmarked small model that delivers on a narrow promise: real agentic capability that runs entirely on a phone or laptop CPU, with day-one support across the major inference stacks. It is not a general-purpose or coding model, and its license carries usage terms worth reading closely, but for developers building local-first agents, assistants, or robotics applications, it is one of the more capable options released so far in its size class.

Editor's Verdict

Liquid AI Ships LFM2.5-2.6B, an On-Device Agent Model brings real, demonstrable value, though with caveats that deserve weighing.

The strongest case for paying attention is fully on-device agentic inference with near-zero marginal cost and strong data-privacy properties, which raises the bar for what readers should now expect from peers in this space. Reinforcing that, broad day-one runtime and hardware support across five inference engines and four hardware vendors adds practical value rather than just headline appeal. On the other side of the ledger, Liquid AI itself excludes agentic coding and knowledge-heavy tasks from the model's recommended use cases is a real constraint, not a marketing footnote, and it should factor into any serious decision. Layered on top of that, the LFM1.0 license imposes usage restrictions rather than a fully permissive open-source license narrows the set of teams for whom this is an obvious yes.

For product teams, content creators, and knowledge workers looking to upgrade a specific workflow, a measured trial makes sense, with clear criteria for when to expand or pull back. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.