The Big Story
In what's being hailed as a groundbreaking achievement, Liquid AI has announced the release of LFM2.5-230M, its smallest model yet. This 230M-parameter, open-weight model is capable of running on-device at an impressive 213 tok/s on a Galaxy S25 Ultra and 42 on a Raspberry Pi 5. Built on the LFM2 architecture, this new model offers unprecedented flexibility and efficiency.
According to the release notes, LFM2.5-230M features support for llama.cpp, MLX, vLLM, SGLang, and ONNX, making it an incredibly versatile tool for developers and researchers alike. With its reduced footprint and increased performance, this model is poised to revolutionize the field of on-device inference.
As reported by MarkTechPost, the LFM2.5-230M release marks a significant milestone in Liquid AI's efforts to democratize access to advanced AI models. With its focus on ease of use and portability, this model has the potential to empower a wide range of users, from hobbyists to professionals.
The implications of LFM2.5-230M are far-reaching, with potential applications in areas such as natural language processing, computer vision, and more. As the AI landscape continues to evolve at breakneck speed, innovations like this will be crucial in driving progress and shaping the future of our industry.
What Shipped
A number of exciting open-source releases and innovations have shipped recently. One notable example is the release of DSpark, a speculative decoding framework from DeepSeek that accelerates per-user generation by 60-85% over MTP-1. According to MarkTechPost, DSpark attaches a draft module to existing DeepSeek-V4 weights and pairs a parallel draft backbone with a lightweight Markov head to cut throughput.
Another significant release is Bashblog, a single bash script that enables the creation of blogs. According to GitHub, this tool provides an easy-to-use framework for bloggers and developers alike.
In addition to these open-source releases, several new AI models have shipped recently. For instance, Liquid AI has released LFM2.5-230M, its smallest model yet, featuring support for llama.cpp, MLX, vLLM, SGLang, and ONNX. As reported by MarkTechPost, this model runs on-device at an impressive 213 tok/s on a Galaxy S25 Ultra and 42 on a Raspberry Pi 5.
From the Labs
Fable 5 Traces Workflow in Colab: Parsing Tool Calls, Auditing Data, and Training Baselines
In this tutorial, we build a stable workflow around the Fable 5 Traces dataset from Hugging Face.
DeepSeek Releases DSpark, a Speculative Decoding Framework That Accelerates DeepSeek-V4 Per-User Generation 60–85% Over MTP-1
According to MarkTechPost, DSpark attaches a draft module to existing DeepSeek-V4 weights and pairs a parallel draft backbone with a lightweight Markov head to cut throughput.
The Fittest Founder in the Room Got Cancer. Here's How He Used AI to Fight Back
When confronted with cancer, Connor Christou fed everything tied to his regime — blood results, scan data, wearable output, journal entries — into Claude.
Other Notable News
SoftBank's CEO isn't the only one with questions about Elon Musk's orbital data center hype, as reported by TechCrunch.
Paul Meade, the Apple vice president in charge of the Vision Pro headset, is reportedly leaving the company to join OpenAI's hardware team, as reported by TechCrunch.
When confronted with cancer, Connor Christou fed everything tied to his regime — blood results, scan data, wearable output, journal entries — into Claude, using AI to fight back, as reported by TechCrunch.
The fittest founder in the room got cancer. Here's how he used AI to fight back, as reported by TechCrunch.
The Take
As we navigate the complex landscape of AI innovation, this week's developments have left us pondering the interplay between stability, efficiency, and progress. The release of LFM2.5-230M by Liquid AI, for instance, marks a significant milestone in the pursuit of on-device inference. This 230M-parameter model, built on the LFM2 architecture, has been shown to run at an impressive 213 tok/s on a Galaxy S25 Ultra and 42 on a Raspberry Pi 5.
Meanwhile, DeepSeek's open-sourcing of DSpark, a speculative decoding framework, is poised to accelerate per-user generation for their DeepSeek-V4 model. This development underscores the importance of continued innovation in AI, as we strive to optimize performance and efficiency.
In related news, SoftBank's CEO has raised questions about Elon Musk's orbital data center hype, highlighting the need for a nuanced understanding of the implications and feasibility of such ambitious projects.
On the personnel front, Apple Vision Pro exec Paul Meade is reportedly leaving to join OpenAI's hardware team, marking an intriguing shift in the AI landscape. As we reflect on these developments, it becomes clear that the pursuit of AI innovation requires not only technological advancements but also the right talent and vision.
As we look ahead, it will be essential to continue exploring the intersection of AI, efficiency, and progress. Will the release of LFM2.5-230M serve as a catalyst for further breakthroughs in on-device inference? How will DeepSeek's DSpark framework shape the future of per-user generation? And what does Paul Meade's move to OpenAI's hardware team mean for the trajectory of AI development?