Liquid AI releases LFM2.5-230M: 213 tokens/s on Galaxy S25 and llama.cpp support
Liquid AI has released LFM2.5-230M, the smallest open-weight model in its lineup. With 230 million parameters, it hits 213 tokens/s on the Galaxy S25 Ultra and 42 on a Raspberry Pi 5, fully offline with no cloud. On instruction following, it outperforms Alibaba's Qwen3.5-0.8B and Google's Gemma 3 1B, which are roughly three to four times larger. The model supports llama.cpp, MLX, vLLM, SGLang and ONNX.
AI-processed from MarkTechPost; edited by Hamidun News
Liquid AI has released LFM2.5-230M, the most compact open-weight model in its lineup. At 230 million parameters, it fits on a smartphone or a single-board computer, and on the tasks it was built for it outperforms competitors with three to four times as many parameters.
What LFM is and how it differs
LFM stands for Liquid Foundation Model, a proprietary architecture from Liquid AI, a company founded by MIT alumni. The approach differs fundamentally from standard transformers: instead of the classic attention mechanism, it uses a hybrid design inspired by neural differential equations. The result is models that run more efficiently with fewer parameters.
LFM2.5-230M is the smallest in the series but is built on the same foundation as the more powerful versions. It does not claim to be a general-purpose assistant: the model is tuned for tool use (calling external tools and APIs in agentic pipelines) and data extraction (structured extraction of data from unstructured text). On exactly these tasks, it delivers better results than significantly larger competitors.
Speed on real hardware and benchmark accuracy
Liquid AI tested performance not on servers but on consumer devices:
- Galaxy S25 Ultra: 213 tokens per second
- Raspberry Pi 5: 42 tokens per second
For context, a comfortable reading speed is about 15–25 tokens/s. On a smartphone, the model runs with roughly eight times that headroom, enough even for real-time interactive applications.
In practice, this means LFM2.5-230M can run offline, without API keys, without cloud costs and without sending data to third-party servers. For enterprise products with confidentiality requirements, that is a weighty argument in itself.
On instruction-following tests, the model beat Alibaba's Qwen3.5-0.8B (more than three times larger) and Google's Gemma 3 1B (four times larger). This is not a win in an overall ranking, but specifically on the tasks the model was designed for.
Supported runtimes
LFM2.5-230M ships with open weights and supports the full standard inference stack:
- llama.cpp: CPU execution without a GPU on any hardware
- MLX: optimized for Apple Silicon chips (M1–M4)
- vLLM and SGLang: for high-load server deployment
- ONNX: a cross-platform standard for production deployment
Coverage is maximal: from a MacBook to a Linux server, from a flagship Samsung phone to a single-board computer costing $80. For open models, breadth of ecosystem support is one of the main factors in real-world adoption.
What this means
Liquid AI shows clearly that architectural efficiency is displacing the parameter race. A 230M-parameter model that runs on a smartphone faster than a person can read, and still beats rivals four times its size, is a convincing argument for specialization over generality. For developers of mobile AI apps and agentic pipelines, this opens a new window of opportunity.
- Nov 1, 2026MoneyGate safety layer blocked all 42 fraudulent requests to local AI models
- Nov 1, 2026ThinkJev Makes AI Think Only About Hard Questions: 28% Fewer Tokens
- Oct 31, 2026Open-source LLM gateway KTMS 1.0 plugs your own GPUs into the cloud and survives outages
- Oct 31, 202622B-Parameter LTX-2.5 Runs on a 10 GB RTX 3080, but Takes 5 Minutes per Shot
Key Questions Answered
What LFM is and how it differs?
LFM stands for Liquid Foundation Model, a proprietary architecture from Liquid AI, a company founded by MIT alumni. The approach differs fundamentally from standard transformers: instead of the classic attention mechanism, it uses a hybrid design inspired by neural differential equations. The result is models that run more efficiently with fewer parameters. LFM2.5-230M is the smallest in the series but is built on the same foundation as the more powerful versions. It does not claim to be a general-purpose assistant: the model is tuned for tool use (calling external tools an
What are the key details on Speed on real hardware and benchmark accuracy?
Liquid AI tested performance not on servers but on consumer devices: Galaxy S25 Ultra: 213 tokens per second Raspberry Pi 5: 42 tokens per second For context, a comfortable reading speed is about 15–25 tokens/s. On a smartphone, the model runs with roughly eight times that headroom, enough even for real-time interactive applications. In practice, this means LFM2.5-230M can run offline, without API keys, without cloud costs and without sending data to third-party servers. For enterprise products with confidentiality requirements, that is a weighty argument in itself. On inst
What are the key details on Supported runtimes?
LFM2.5-230M ships with open weights and supports the full standard inference stack: llama.cpp: CPU execution without a GPU on any hardware MLX: optimized for Apple Silicon chips (M1–M4) vLLM and SGLang: for high-load server deployment ONNX: a cross-platform standard for production deployment Coverage is maximal: from a MacBook to a Linux server, from a flagship Samsung phone to a single-board computer costing $80. For open models, breadth of ecosystem support is one of the main factors in real-world adoption.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.