Liquid AI's LFM2.5-2.6B Runs AI Agents on a Raspberry Pi

Liquid AI's new LFM2.5-2.6B is a 2.6-billion-parameter open-weight model that runs agent workloads on hardware as modest as a Raspberry Pi — no cloud, no GPU required. For startup founders and enterprise teams navigating privacy regulations or edge deployments, it reframes the cost-versus-capability tradeoff for small models in ways worth understanding.

Liquid AI Releases a 2.6B-Parameter Agent Model That Runs on a Raspberry Pi

Liquid AI, the MIT-spinout founded in 2023, this week released LFM2.5-2.6B, an open-weight language model built from the ground up for agentic tasks on local hardware. The headline claim is striking: meaningful AI agent performance — tool calling, document management, calendar automation, always-on background workflows — running on devices as small as a Raspberry Pi, with no cloud dependency and no GPU required.

That's not just a spec sheet flex. It represents a genuine architectural bet that the next wave of enterprise AI adoption will be won not at the frontier but at the edge.

Why This Matters Beyond the Benchmarks

The conventional framing in AI model releases is a race to the top — bigger context windows, better MMLU scores, stronger coding benchmarks. Liquid is making a deliberately different argument: a sufficiently capable small model unlocks entirely new deployment categories where latency, data privacy, connectivity, and inference cost matter more than leaderboard position.

"You should use [edge AI] when you can't use a cloud model," said Maxime Labonne, Liquid AI's head of post-training, in an interview with VentureBeat.

For enterprise teams in healthcare, finance, legal, or defense — where sensitive data simply cannot leave the building — this isn't a nice-to-have. It's a prerequisite. And for startups building products in connectivity-constrained environments like vehicles, robotics, or industrial IoT, a model that fits in under 2.5 GB of RAM and runs at 30 tokens per second on a smartphone is a different kind of unlock than anything running on a cloud API.

The Specs

2.6 billion parameters, 128K-token context window, native tool calling
Runs on CPU; ~220 tokens/sec on Apple M5 Max, ~113 tokens/sec on AMD Ryzen AI Max+ 395, ~30 tokens/sec on a smartphone
Under 2.5 GB memory footprint
On a single Nvidia H100: ~15,000 output tokens/second under concurrent load (~1.3B tokens/day)
Available on Hugging Face with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX
Base checkpoint also available for fine-tuning; open-source fine-tuning framework LEAP included

(Vendor-reported figures; not independently verified.)

Built for Agents, Not Chatbots

LFM2.5-2.6B was pretrained on approximately 34 trillion tokens and trained specifically inside production agent harnesses — including Hermes Agent and OpenClaw — rather than evaluated primarily against conversational benchmarks. Liquid doubled the vocabulary to 128K tokens to better support non-Latin scripts and extended the context window mid-training for long-running workflows.

The post-training pipeline is notably sophisticated: supervised fine-tuning → teacher specialization (separate expert models for instruction following, math, code, and tool use) → multi-domain on-policy distillation (MOPD) to recombine those specialists → agentic reinforcement learning inside real production harnesses.

Labonne described a welcome side effect: "We got better at math, at instruction following... we even got really good at code" — gains that extended well beyond the original agentic targets.

What Founders and Builders Should Watch

The licensing fine print. Like Moonshot AI's Kimi K3 released last month, LFM2.5-2.6B ships under a custom open-weights license rather than a standard permissive one (MIT, Apache 2.0). Enterprise legal teams should review usage terms before building products on top of it — "open weights" and "open source" are not synonymous.

The agentic-first design philosophy is a signal. Liquid is not alone in betting that agents are the primary consumption pattern for LLMs going forward. Building a model evaluated and trained inside agent harnesses rather than on academic benchmarks alone reflects where production deployments are actually headed.

The cost math. Running capable agents at essentially the cost of electricity — on hardware you already own — changes the unit economics of AI-powered features for startups operating under tight margins or in markets where cloud API costs compound fast.

Liquid AI's mobile app Apollo lets users try the model directly on their phones.