Editorial illustration for Liquid AI releases LFM2.5-230M, adds llama.cpp, MLX, vLLM, SGLang, ONNX
Liquid AI LFM2.5-230M Adds 5 Inference Engine Support
Liquid AI releases LFM2.5-230M, adds llama.cpp, MLX, vLLM, SGLang, ONNX
Liquid AI just released LFM2.5-230M. It is a small model, 230 million parameters, but it runs on almost anything: llama.cpp, MLX, vLLM, SGLang, ONNX. The benchmarks look like a misprint.
It beats Qwen3.5-0.8B and Gemma 3 1B IT on instruction following and data extraction. That is a 230M model outperforming models four times its size. The secret is distillation, the smaller model inherits behavior from the 350M sibling, then gets fine-tuned in three stages.
Honesty comes with the package: Liquid AI says don’t use it for reasoning, math, or code. But for large-scale data extraction, parsing 100,000 clinical reports into structured fields, this is exactly the right tool. And now it ships with support across the major inference engines, ready for on-device deployment.
Liquid AI shipped LFM2.5-230M , it’s the company’s smallest model to date. The release targets a specific job: running agentic tasks on phones, robots, and automation devices.
This is a model that knows exactly what it is, and what it isn’t. LFM2.5-230M doesn’t pretend to be a universal oracle. It’s a scalpel, not a sledgehammer.
On instruction following and structured data extraction, it punches far above its weight class, besting models several times its size. That’s the distillation advantage: targeted behavior inherited from a bigger sibling, without the bloat. And the framework support?
That’s the quiet revolution here. llama.cpp, MLX, vLLM, SGLang, ONNX, pick your stack. Deploy it on a phone, a laptop, or a low-power edge device.
The inference efficiency is built in, not bolted on. So don’t ask what this model can’t do. Ask what you actually need it to do.
If the answer is parse a million clinical reports, follow complex instructions, or extract structured fields from messy text, this is the tool. It’s honest about its limits. That honesty is the real strength.
In a field obsessed with bigger, faster, broader, Liquid AI built something useful precisely because it stayed small and focused. That’s not a compromise. That’s strategy.
Related Reading
Browse all Meta news →
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Related Topic OpenAI hires 630 ex-Meta staff as ChatGPT memory may turn data into ads
Business & Startups
Related Topic Meta AI Update Pulls From Your Calendar for Daily Briefings
LLMs & Generative AI
From Archives Meta's Astryx adds CLI and MCP server to design system used by Figma, Snowflake
From Archives Apple Vision Pro exec departs for OpenAI as Apple eyes cheaper glasses vs Meta
From Archives Calibration uses NVIDIA Triton Llama-3-8B A10 and vLLM Qwen2.5-7B RTX 4090 data
Research & Benchmarks
From Archives Meta says AI moderators make 13% fewer errors than humans, defends rollout speed
LLMs & Generative AI
Common Questions Answered
What specific model did Liquid AI release with the LFM2.5-230M designation?
Liquid AI released the LFM2.5-230M, a 230 million parameter language model that builds on their prior LFM2.5 architecture. This model is optimized for edge deployment and offers competitive performance despite its relatively small size.
Which inference frameworks are now supported by Liquid AI's LFM2.5-230M?
The LFM2.5-230M now supports llama.cpp, MLX, vLLM, SGLang, and ONNX frameworks. This broad compatibility allows developers to run the model on CPUs, Apple Silicon, and through high-performance serving systems.
What advantage does adding llama.cpp support provide for the LFM2.5-230M?
llama.cpp support enables efficient CPU inference, making the LFM2.5-230M accessible on a wide range of hardware without requiring a GPU. This integration is particularly useful for on-device applications and local deployments in resource-constrained environments.
How does vLLM integration benefit users deploying Liquid AI's LFM2.5-230M?
vLLM integration allows users to serve the LFM2.5-230M with high throughput and low latency using advanced batching and memory management techniques. This makes the model suitable for production-scale applications requiring real-time responses.
Further Reading
- Liquid AI releases LFM2.5-230M, a 230-million-parameter powerhouse for edge devices — Digg
- LFM2.5-230M: Built to Run Anywhere — Liquid AI
- Liquid AI's smallest model yet LFM2.5-230M beats models 4X its size at data extraction, can run anywhere — VentureBeat
- LiquidAI/LFM2.5-230M - vLLM Recipes — vLLM
- LiquidAI/LFM2.5-230M - Hugging Face — Hugging Face
