Liquid AI's 2.6B model runs agents on your phone, beats models 4x its size

Liquid AI just shipped a 2.69-billion-parameter model that doesn’t just chat—it plans, uses tools, and executes multi-step tasks entirely on-device. No cloud. No API calls. No per-token billing. We’re talking phones, laptops, robots, and air-gapped servers running agentic workloads locally. The marginal cost per inference round is effectively zero, and your data never leaves the hardware. That’s a fundamentally different value proposition from the subscription-plus-cloud model most AI companies are pushing.

The numbers are striking for a model this lean. LFM2.5-2.6B decodes at 220 tokens per second on an M5 Max while consuming under 2.5 GB of memory. On a phone, you’re still getting roughly 30 tokens per second. Liquid AI reports that a single NVIDIA H100 SXM5 can serve about 1.3 billion tokens per day. The pre-training budget was approximately 34 trillion tokens across a 128,000-token vocabulary, and the architecture mixes 22 double-gated short convolution blocks with 8 grouped-query attention blocks. It’s a hybrid design that prioritizes efficiency without fully abandoning the attention mechanisms the industry knows well.

What makes this model genuinely unusual is the four-stage post-training pipeline that turns the base checkpoint into an agent. Two rounds of supervised fine-tuning use a dataset roughly seven times larger than what went into the company’s earlier 8-billion-parameter model. Then comes teacher specialization with reinforcement learning, followed by on-policy distillation where a student model rolls out under its own policy and routes prompts to domain-specific teachers. The final stage is agentic RL inside real harnesses like Hermes Agent and OpenClaw. This isn’t a generic instruct model with a system prompt slapped on top—it was purpose-built for tool calls and multi-step workflows.

Benchmark results back up the ambition. LFM2.5-2.6B leads every instruction-following test Liquid AI published and nearly every tool-use benchmark, beating Google’s gemma-4-E4B-it (8B parameters) and Qwen3.5-9B on ToolSandbox, Multi-IF, and IFStruct. The one place larger models still dominate is coding: LiveCodeBenchv6 shows Qwen3.5-9B at 69.86 versus 59.41. Liquid AI is refreshingly upfront about this gap and explicitly recommends against using the model for agentic coding or knowledge-heavy tasks. Instead, the target use cases—on-device assistants, offline document triage over 128K-token inputs, form extraction, robotics command parsing—play to the model’s strengths in tool orchestration and constrained environments. This is a pragmatic bet on a world where latency, privacy, and cost matter more than raw knowledge recall.