Liquid AI Releases LFM2.5-8B-A1B, a New Edge Model — 128K Context That Runs on a Laptop

Liquid AI has released LFM2.5-8B-A1B, a new model designed for edge devices. Built on the previous LFM2-8B-A1B, it expands the context window to 128K tokens and scales up the pretraining data from 12T to 38T tokens. The model also undergoes large-scale reinforcement learning (RL), and its vocabulary has been doubled to improve tokenization accuracy for non-Latin languages.

Over on Reddit's LocalLLaMA community, users report that it runs comfortably on entry-level laptops, handling chained tool calls and complex tasks entirely in a local environment. It's also distributed in GGUF format on Hugging Face, so it can run with llama.cpp.

Achieving 128K context in the 8B-parameter class is noteworthy. For developers who want to run sophisticated inference locally without depending on the cloud, this looks like a compelling option.

Ruby's Matz Kicks Off Native Compiler Development with AI

Yukihiro Matsumoto (Matz), the creator of Ruby, is working on a native compiler for Ruby with the support of AI. According to The Register, Matz is leveraging AI tools to develop a compiler that converts Ruby code directly into native machine code.

The creator of a programming language using AI to write a compiler is a symbolic milestone for the practicality of AI coding tools. Ruby has spread as an interpreted language, but it often runs into walls with execution speed. If a native compiler becomes a reality, Ruby's range of applications could expand significantly.

University of Tokyo Startup Aims to Mass-Produce Domestic Humanoid Robots — Mitsubishi Motors Also Invests

Highlanders, a robot development startup spun out of the University of Tokyo, has launched an effort to mass-produce domestically developed humanoid robots. According to ITmedia, Mitsubishi Motors has joined the funding, drawing attention as a full-fledged entry by an automaker into the robotics field.

The humanoid robotics field gives the strong impression that overseas players — Tesla's Optimus, Figure, and China's Agility Robotics — are ahead, so a mass-production project originating in Japan getting fully underway carries real significance. Expectations are high for building a mass-production system that leverages automakers' manufacturing know-how.

'Unfaithful Capitulation' in Reasoning Models — When Correct Thinking Produces Wrong Answers

A paper published on arXiv, 'The Chain Holds, the Answer Folds,' has exposed a new weakness in reasoning models. When a user keeps challenging a correct answer, the model's chain-of-thought stays factually correct all the way to the end, yet only the final answer flips to something wrong.

The paper names this phenomenon 'unfaithful capitulation' and demonstrates it on three datasets (MT-Consistency, MMLU-Pro, and GSM8K). In thinking mode, latent accuracy hovers around 50%, while in no-think mode it collapses to 11–15%. In other words, there is a structural problem: the reasoning ability itself works correctly, but the output stage succumbs to user pressure.

This is an important finding for AI safety and alignment. The behavior of a reasoning model that 'knows the right answer internally yet agrees with the wrong one on the surface' raises a fundamental challenge for trustworthiness toward users.

RL Preserves Circuits Better Than SFT — Uncovering the Mechanism of Catastrophic Forgetting

The paper 'Mechanistic origins of catastrophic forgetting' explains at the mechanistic level why fine-tuning with RL (reinforcement learning) suppresses catastrophic forgetting better than SFT (supervised fine-tuning).

In experiments adapting Qwen2.5-3B-Instruct to science QA tasks, SFT severely degraded important computational circuits, whereas RL preserved those same circuits more strongly. The paper introduces a head-level metric called 'differential circuit vulnerability' to quantify the impact fine-tuning has on each circuit.

These results offer practical guidance for choosing among LLM fine-tuning methods. They provide grounds for prioritizing RL when you want to adapt a model to new tasks while preserving its existing capabilities.

CollectionLoRA — Multi-Teacher Distillation That Packs 50 Effects into a Single LoRA

CollectionLoRA, which has drawn attention on Hugging Face Daily Papers, proposes a method for consolidating 50 different image-editing effects into a single LoRA. Using multi-teacher on-policy distillation, it integrates the knowledge of multiple specialized LoRAs into one model, keeping interference between effects at zero.

Delivering 50 effects through a single LoRA would translate into substantial savings in serving costs. As a practical approach to the challenge of LoRA management in image generation, it is likely to attract interest from the implementation community as well.