Liquid AI LFM2.5 Brings Local AI Agents to Raspberry Pi
Liquid AI’s 2.6B-parameter LFM2.5 runs agents on phones and Raspberry Pi devices, challenging larger models but imposing a $10M revenue license cutoff.
Summary
On August 6, 2026, VentureBeat reported that Liquid AI, founded in 2023 by former MIT computer scientists, released LFM2.5-2.6B, a 2.6-billion-parameter, open-weight, text-only dense model for private, low-cost local agents. Its 128,000-token context, native tool calling, post-trained and Base checkpoints are on Hugging Face with llama.cpp, MLX, vLLM, SGLang and ONNX support; LEAP enables open-source fine-tuning. It runs on CPUs from phones to Raspberry Pi devices without cloud inference or GPUs, under 2.5 GB. Unverified Liquid benchmarks show 30 tokens per second via Apollo, 113 on AMD Ryzen AI Max+ 395, 220 on Apple M5 Max, and nearly 15,000, or 1.3 billion daily, on one Nvidia H100.
Pretraining used about 34 trillion tokens, a 128K vocabulary supporting non-Latin scripts and mid-training for 128K context. Post-training joined supervised fine-tuning, specialist teachers, multi-domain on-policy distillation and reinforcement learning inside Hermes Agent and OpenClaw on research, coding, documents, tools and automation. Liquid built a phone-native harness for proactive background agents; Pi and OpenAI-compatible endpoints also work. Maxime Labonne recommends the model for high-volume tools, calendars, workflows, vehicles and robotics, not coding-heavy work, and says fine-tuning can match GPT and Claude on less-complex tasks.
Liquid’s charts beat Google Gemma 4 E2B, 5.1B, and E4B, 8B, plus Alibaba Qwen3.5-4B, 4.7B, and 9B, 9.7B, on instruction following and most tool use: ToolSandbox scored 77.83 versus Qwen 9B’s 76.44; BrowseComp+ 26.89 versus 27.23. Atomic Chat said 35 calls across three tasks ran 3.7 times faster than 284B DeepSeek-V4-Flash, while Qwen leads math and coding. Apache 2.0 covers Gemma and Qwen, MIT covers DeepSeek; LFM Open License v1.0 permits commercial use below $10 million annual revenue, requiring larger companies to negotiate. MacPaw will put custom Liquid models in Eney on Apple silicon through Elix and Mnemos, with results due later in 2026.
Positives
- Under 2.5 GB of memory lets LFM2.5-2.6B run locally on smartphones, laptops and Raspberry Pi devices without GPUs or cloud inference.
- 128,000 tokens of context and native tool calling support long-running document, calendar and workflow agents.
- 35 tool calls across three Atomic Chat tasks finished 3.7 times faster than the 284B-parameter DeepSeek-V4-Flash.
- 77.83 on ToolSandbox exceeded Qwen3.5-9B’s 76.44 despite LFM2.5-2.6B having nearly four times fewer parameters.
- MacPaw plans to integrate custom Liquid models into Eney through its Elix inference engine and Mnemos memory layer later in 2026.
Risks & concerns
- $10 million in annual revenue triggers a separate commercial agreement, unlike the permissive Apache 2.0 and MIT licenses covering major rivals.
- Nearly 15,000 output tokens per second on one Nvidia H100 and Liquid’s other performance figures have not been independently verified.
- Text-only LFM2.5-2.6B requires separate Liquid models for vision and audio, while Gemma 4 and Qwen3.5 offer native multimodality.
- Qwen3.5 retains stronger math and coding performance, including leadership on AIME25 and LiveCodeBench.
- Coding-heavy and highly complex work remains better suited to larger cloud models despite Liquid’s post-training improvements.