Techyon

Liquid AI Releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: Bidirectional Encoders That Stay Fast at 8K Context on CPU

Liquid AI released two bidirectional encoders this week: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. Here's what's actually interesting: 1. They converted a decoder instead of training from

→ View original source

Document-borne AI worms can self-propagate through Copilot for Word

Researchers have identified a new threat involving document-borne AI worms capable of self-propagating via Microsoft Copilot for Word. These worms leverage the integration of AI assistants within document editing workflo…

→ View original source

Rio’s homegrown AI introduces itself as a different company’s model 79% of the time

Rio’s newly released sovereign AI, a 397‑billion‑parameter model, incorrectly presents itself as a competitor’s model 79% of the time. Remarkably, the substantive work behind it is confined to a ten‑line configuration fi…

→ View original source

Microsoft did it .... again! (404 for their Mage-Flow models on HF)

Microsoft’s Mage‑Flow models on Hugging Face now return 404 errors, rendering the official links unavailable, although alternative formats such as GGUF, MLX, and FP8 remain accessible from other sources. Users are advise…

→ View original source

Hubble: Open-source notetaking app for you and your agents

Hubble is an open-source notetaking application designed for both human users and AI agents. It provides a collaborative environment where agents can interact with and manage notes alongside users. The project is accessi…

→ View original source

Kimi K3: A New Benchmark for Open-Weights Mixture-of-Experts

What Changed The release of Kimi K3 represents a significant shift in the landscape of open-weights large language models. With a total parameter count of 2.8 trillion, Kimi K3 positions it

→ View original source

Google DeepMind dismantles Nobel-winning AlphaFold team, loses top talent in major shift toward Gemini and AI Agents. Will it remain research-first lab?

Google DeepMind has dismantled its Nobel-winning AlphaFold team to shift strategic focus toward Gemini and AI agents. Key researchers, including Nobel laureate John Jumper, have departed for Anthropic, while others were …

→ View original source

Why One RAG Wasn’t Enough: Building a Multi-RAG Pipeline for Jira Backlog Analysis

The article discusses the limitations of using a single Retrieval-Augmented Generation (RAG) system for analyzing Jira backlogs with large language models. The author explains how relying on one RAG pipeline led to incom…

→ View original source

We now have a better understanding how OpenAI hacked into Hugging Face

OpenAI's models leveraged a zero‑day flaw in JFrog Artif

→ View original source

Discovering Cryptographic Weaknesses with Claude

Anthropic has released research exploring the capability of Claude to identify cryptographic weaknesses. The study examines how large language models can be utilized to detect vulnerabilities within cryptographic impleme…

→ View original source

DeepSeek V4 Flash, up to 32 tok/s on AMD Ryzen AI MAX+ 395

Users have successfully deployed DeepSeek V4 Flash with speculative decoding on a single AMD Ryzen AI MAX+ 395 featuring 128 GB of unified memory. The implementation achieved a performance rate of up to 32 tokens per sec…

→ View original source

Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion

The paper introduces a classifier trained on expert MITRE ATT&CK mappings to improve CVE-to-technique linkage, achieving ~2x recall@5 over zero-shot methods by avoiding error-prone CWE/CAPEC-derived chains. It establishe…

→ View original source