Liquid AI's LFM2.5-2.6B Beats 9B Models Running Entirely on Your Phone
Liquid AI's LFM2.5-2.6B runs full agentic workflows on-device, beating models nearly 4x its size on tool use and instruction following benchmarks.
- New release: LFM2.5-2.6B is a 2.6B parameter agentic model designed to run fully on-device with no cloud dependency.
- Benchmark performance: Beats Qwen3.5-9B on ToolSandbox (77.83 vs 76.44) and all instruction-following benchmarks despite being 4x smaller.
- Speed: 220 tok/s on Apple M5 Max, 113 tok/s on AMD Ryzen CPU, 30 tok/s on a phone, all under 2.5 GB memory.
- Architecture: Hybrid LFM2 design with 22 short convolution blocks + 8 attention layers, trained on 34T tokens with 128K context.
- Training: Four-stage post-training pipeline including agentic RL inside real harnesses (Hermes Agent, OpenClaw, Pi) using GRPO.
- Limitations: Not recommended for coding-heavy or knowledge-intensive agentic tasks; larger models still hold an edge there.
The premise of on-device AI agents has always been compelling but frustratingly out of reach. Models small enough to run locally tend to fall apart the moment you ask them to plan across multiple steps, call tools reliably, or follow complex instructions. Liquid AI is making a direct push against that wall with LFM2.5-2.6B, a 2.6B parameter model purpose-built for agentic workloads that runs entirely on-device, no cloud calls required.
The core pitch: zero marginal cost agents
Unlike agents that depend on cloud APIs, local agents give you free inference, low latency, and real privacy. Removing the per-token cost changes how developers build: agents can now be massively parallelized on local hardware, running background tasks that burn through millions of tokens at no marginal cost. That's not just a cost argument. It's an architectural unlock. When token spend is no longer a constraint, you can run agents continuously in the background, on the user's own hardware, across any number of parallel tasks.
LFM2.5-2.6B is small enough to run on a phone, fast enough to stay responsive on a CPU, and capable enough to power agentic workflows: planning, calling tools, and tackling multi-step tasks.
The architecture under the hood
LFM2.5-2.6B is built on the LFM2 hybrid architecture, which is the key reason it punches above its weight on constrained hardware. LFM2 pairs efficient gated short convolution blocks with a small number of grouped query attention (GQA) blocks. In practice, attention blocks account for only about 37% of the model. The remaining 63% is made up of cheaper convolution blocks.
Why does that ratio matter? In a traditional Transformer, every layer uses Softmax Attention, which scales quadratically with sequence length. This leads to massive KV caches that devour VRAM. The LFM2 convolution blocks sidestep this entirely. Each convolution layer maintains a fixed-size state that gets updated on each new token. There is no attention over past tokens and no KV cache: the state memory at token 10,000 is the same as at token 10.
The specific block type is called an LIV (Linear Input-Varying) convolution. As the name suggests, it is a linear-time convolution operator whose behavior changes based on the input. Liquid AI ran hardware-in-the-loop architecture search to find the best layer mix across quality, latency, and memory. The search space included GQA, short convolutions, linear attention, S4, Mamba, and Mamba2. The winner was the simplest option: mostly short convolutions, a handful of attention layers. Once a handful of GQA blocks are available to handle long-range retrieval, the inexpensive gated short convolution alone is sufficient to reach the best quality-latency-memory trade-off.