Researchers from Alibaba and several Asian universities built Co-Design Bench from real user change requests, requiring coding agents to rework existing web apps with multimodal inputs such as images, audio, video and documents. The best of ten tested models, GPT-6 Astra, met only about 70 percent of requirements on average and fully completed roughly one in four tasks, sharply contrasting with near-perfect scores on older benchmarks.
Multimodal Input
多模态输入
The ability of a model to simultaneously receive diverse data types such as text, images, or audio.
Curated Intelligence
Amazon is adding a text-message interface to Alexa for Shopping, letting users describe or photograph products and have the assistant search listings, reply with recommendations, and place orders on their behalf. The feature extends the shopping assistant launched in May, targeting lighter-weight ordering moments and adding preference memory and price-drop alerts.
Open-source tool Penmentor integrates LLMs like Claude to provide one-click handwritten math grading in xournal++. It simulates a human tutor by circling errors, offering hints instead of answers, and generating Manim-style animations for conceptual clarity.
Liquid AI released its new decision model, d1-ultra, claiming superior performance across eight task categories compared to TypeSafe AI's recently launched Jev 1.14. While Jev 1.14 introduced image input support, longer context windows, and a new 'box' type, the rapid succession of d1-ultra highlights the intense competition in decision-making models. However, current claims rely heavily on vendor self-reported benchmarks without independent verification or detailed technical reports.
Meta has released Muse Spark Open, an 816B-parameter (42B active) open-weights model that scores 39 on the Artificial Analysis Intelligence Index, surpassing NVIDIA's Nemotron to become the top US open-weights model. The model supports multimodal inputs (text, image, video, audio), offers high token efficiency with only 40k output tokens per task, and demonstrates reduced hallucination rates alongside improved performance on automation benchmarks. Weights are available on Hugging Face, with API access provided through major inference platforms.
Jev 1.14 has been released, featuring image multimodal inputs, a new Box API, improved generation quality, and extended context length. Databricks announced a partnership with TypeSafe to integrate this version for enterprise customers. The team highlights that this release targets early adopters via an opt-in 'danger zone' switch, distinguishing it from the stability-focused infrastructure versions.
MiniMax has released FastH3-Omni, an open-source model claiming faster-than-realtime video generation. It supports text, audio, and image inputs, currently accessible primarily via API for developers.
Inception Labs' Mercury Decide model now supports image input via OpenRouter, enabling users to combine visual and textual data in decision requests. The update allows the model to answer multiple-choice questions, provide scores, or output probabilistic yes/no judgments based on images, with pricing remaining unchanged at free endpoints or $0.02 per million input tokens.
ShipStation Global developed 'Deckhand,' an AI-powered IT help desk agent for Slack built on Amazon Bedrock AgentCore. By integrating curated knowledge bases, Jira Service Management, and multimodal screenshot analysis, the agent achieved a 97.5% automatic resolution rate and 96% positive employee feedback.
Tech commentator Greg Isenberg highlights that static 2021 internet data, lack of long-term memory, inability to process multimodal inputs, and text-only responses defined early Large Language Models. He argues these current limitations will be viewed by future generations as 'Stone Age' bottlenecks, emphasizing the historical discontinuity between pure-text static models and evolving agents with real-time perception and memory.