LFM2.5-2.6B Runs Tool-Calling AI Agents in 2.5GB of RAM

A 2.6B Model That Behaves Like an Agent, Not Just a Chatbot

Liquid AI released LFM2.5-2.6B on August 4, 2026, and the interesting claim is not the parameter count. It is what the model is built to do at that size. This is a 2.69 billion parameter model trained specifically for agent workloads — calling tools, following multi-step plans, and running structured workflows — and it does that in under 2.5GB of memory, which puts it comfortably inside the budget of a phone, a laptop, or a Raspberry Pi with no GPU attached.

• LFM2.5-2.6B has 2.69B total parameters across 30 layers, pre-trained on roughly 34 trillion tokens covering 16 languages
• Context window is 131,072 tokens with a 128,000-token vocabulary
• Reported throughput is 220 tokens per second on an Apple M5 Max and around 30 tokens per second on a smartphone, in under 2.5GB of memory
• Open weights ship under the lfm1.0 license in native, GGUF, MLX, and ONNX formats, with day-one support for llama.cpp, vLLM, SGLang, and LM Studio

Why On-Device Agents Are Harder Than On-Device Chat

Running a small language model locally has been straightforward for a while. Running an agent locally has not. An agent has to emit well-formed tool calls, keep track of state across several turns, recover when a tool returns something unexpected, and follow instructions precisely enough that the surrounding harness can trust its output. Small models have historically been fluent but sloppy at exactly those things, which is why local deployments tended to stop at summarisation and autocomplete.

The architecture here is unusual and clearly aimed at that gap: 22 double-gated short convolution blocks paired with only 8 grouped-query attention blocks. Convolution-heavy designs are cheaper per token than full attention stacks, which is how the model sustains high throughput on hardware with no dedicated accelerator, while the attention blocks that remain carry the long-range reasoning.

How Does It Compare to Larger Models on Tool Use?

Liquid AI's reported benchmarks put LFM2.5-2.6B ahead of models several times its size on agent-shaped evaluations. On ToolSandbox it scores 77.83 against 76.44 for Qwen3.5-9B and 65.00 for gemma-4-E4B-it. On Multi-IF, which measures multi-turn instruction following, it reports 80.07 against 62.55 for the 9B comparison. On IFStruct, a structured-output test, it reports 85.49 against 78.50.

These are vendor-reported figures and should be read as such until independent evaluations land. But the pattern is consistent and it matches the design intent: the model is not trying to win on general knowledge, where a 2.6B model will always lose. It is trying to win on the narrow set of behaviours an agent harness actually depends on, and it appears to have traded capacity toward exactly those.

What Runs It Today

The release is unusually well-provisioned for day-one deployment. Weights are on Hugging Face in native, GGUF, MLX, and ONNX formats, and the supported runtime list covers llama.cpp, vLLM, SGLang, and LM Studio across AMD, Qualcomm, and Apple silicon. If you already have a local inference setup, there is a good chance it works without changes — and if you are still choosing one, our comparison of Ollama, vLLM, and llama.cpp covers the tradeoffs.

For anyone who has been following how far single-board computers can push local inference, this is a meaningful step up from the microcontroller-class experiments. We recently covered an ESP32-S3 running a 28.9M-parameter model at 9 tokens per second, which is an impressive stunt at that scale. A 2.6B agent model at 30 tokens per second on a phone is a different category — it is fast enough to sit inside a real product loop.

Where This Actually Matters

The obvious win is privacy and regulatory posture. Healthcare, legal, and financial workflows frequently cannot send prompts to a cloud endpoint, and an agent that runs entirely on the device removes that objection rather than mitigating it. The less obvious win is cost predictability: a local agent has a fixed hardware cost and no per-token bill, which changes the economics of anything that runs continuously in the background rather than on demand.

The single-H100 figure Liquid AI cites — roughly 1.3 billion tokens served per day — also matters for teams who want to self-host centrally rather than on endpoints. That is a lot of throughput from one accelerator.

More on where edge inference is heading in our AI coverage.

Sources: Liquid AI — August 4, 2026; MarkTechPost — August 6, 2026; VentureBeat — August 2026.