Deploy local agents everywhere with LFM2.5-2.6B

While the industry is still marveling at the power of massive models — and paying sky-high cloud bills — a counterintuitive shift is underway: fitting capable AI agents into your pocket. LiquidAI’s freshly released LFM2.5-2.6B proves that you don’t need a giant model to have a powerful local agent. With only 2.6B parameters, it outperforms Gemma models up to 4x larger on tool use, multi-step reasoning, and instruction following, often matching 9.7B Qwen models.

Why now: the push for local agents The past year saw AI agents go mainstream, but almost all rely on cloud-hosted large models. This brings three pain points: sensitive data must leave the device, latency is unavoidable, and costs scale linearly with usage. What if agents could run natively on users’ phones, laptops, or even IoT devices? Small models have long been the bottleneck — LFM2.5 aims to break that.

How a small model becomes an Agent The secret isn’t architecture; it’s post-training. The team designed a four-stage pipeline, with the final “Agentic RL” step being the true differentiator.

Typical post-training relies on SFT and preference alignment. But to be an agent, a model must learn to plan, call external tools, and adapt based on environmental feedback. LFM2.5’s recipe: first, two rounds of agent-heavy SFT teach basics like tool use and web search. Then, specialist teacher models are trained separately for math, code, tool use, etc. Next, multi-domain on-policy distillation merges them into one 2.6B student. Finally, Agentic RL places the student directly inside real agent harnesses (e.g., OpenClaw, Hermes Agent), runs multi-turn tasks in a sandbox, and provides rewards based on task completion and correct tool usage. The model updates its policy accordingly.

The clever twist: instead of designing synthetic training tasks, they repurpose standard agent evaluation harnesses as the training ground. The model literally learns in the same environment where it will be tested. A “Harness Proxy” captures token-level trajectories for full replay and validation, making the RL loop reliable.

A bigger trend: small models, specialized for agency This reveals a deeper pattern: large models excel at generality, but small models can become specialists through focused training. Agency isn’t a linear function of parameter count — it depends on data and objective alignment. We used to equate “intelligence” with massive compute, but LFM2.5 shows that if the task is clearly scoped to tool calling and multi-step coordination, 2.6B parameters are more than enough. This could catalyze a wave of “local-first agents,” where developers no longer treat privacy and cost as optional trade-offs.

Practical takeaway: what this means for your app If you’re building assistants that involve tool calls, information retrieval, form filling, and similar workflows — and your users care about data privacy — LFM2.5 offers a near-zero-cost local solution. Its inference speed enables real-time responsiveness even offline. Imagine: intelligent photo organization, local file automation, privacy-sensitive calendar management; all now possible without a network connection and without hefty cloud bills.

Surprising insight: big models aren’t a prerequisite for agents Many assume “agent = large model + tools.” LFM2.5 disproves this. On the BFCLv4 tool-use benchmark, it’s narrowly edged out only by a 9.7B Qwen, and it tops several instruction-following tests outright. For structured tool-use tasks, model size isn’t the deciding factor — training strategy is. You probably don’t need a 7B+ model for a local customer service bot; 2.6B suffices, saving significant inference cost.