Liquid AI releases LFM2.5-230M: non-transformer model beats larger rivals at 42 tokens per second on a Raspberry Pi

Liquid AI on June 25, 2026 released LFM2.5-230M, its smallest model yet — a 230-million-parameter non-transformer model that runs at 42 tokens per second on a Raspberry Pi 5 and outperforms models twice to four times its size on data extraction and tool-use benchmarks.

What's new

LFM2.5-230M is built on Liquid AI's Liquid Foundation Model architecture, a non-transformer design based on state-space and liquid neural network continuous-time formulations developed at MIT. This is the smallest model in the LFM2.5 family.

Key specs and performance:

• Parameters: 230 million
• Training: Pre-trained on 19 trillion tokens
• Context window: 32k tokens
• Hardware performance:
  • Samsung Galaxy S25 Ultra: 213 tok/s decode
  • Raspberry Pi 5: 42 tok/s decode (523 tok/s prefill), using only 293 MB memory
• Benchmarks:
  • BFCLv3 tool-use: 43.26 (vs IBM Granite 4.0-350M at 39.58, Google Gemma 3 1B at 16.61)
  • GPQA Diamond: 25.41
  • IFEval instruction following: 71.71

The model is available for download and fine-tuning, described by Liquid AI as "a fast, lightweight foundation for developers to fine-tune and deploy in agentic workflows."

One explicit limitation: the model is "not recommended for reasoning-heavy workloads such as advanced math, code generation, or creative writing" — it's optimized for tool use, data extraction, and instruction following.

Context

Liquid AI, an MIT spinout founded in 2023 and valued at roughly $2 billion, has built its identity around non-transformer architectures. The LFM family uses state-space dynamics and continuous-time neural ODEs rather than the standard attention mechanism.

At 230M parameters, LFM2.5-230M targets the edge tier — IoT devices, phones, and single-board computers where models typically top out at 1B parameters due to memory constraints. Running at 293 MB with competitive tool-use benchmark performance puts it in a practical range for on-device AI pipelines.

The BFCLv3 performance is notable: the benchmark tests function/tool calling, which underpins agentic AI applications. Outperforming IBM's Granite 4.0-350M and Google's Gemma 3 1B-IT on this benchmark at one-quarter the parameter count is a meaningful efficiency claim.

Why it matters

A model that runs on a Raspberry Pi at practical speeds (42 tok/s is fast enough for real-time applications) while maintaining competitive tool-use performance opens up scenarios that weren't viable before — local automation agents, offline processing, privacy-preserving enterprise workflows.

For Liquid AI, LFM2.5-230M gives developers a low-risk entry point to evaluate non-transformer architectures. If the efficiency advantage holds in production, it adds pressure on the ecosystem to look beyond transformer variants for efficiency-sensitive workloads.