Tencent Hunyuan Hy3 Tops OpenRouter LLM Leaderboard
The Hy3 model claims the number one spot, marking a milestone for Tencent's AI lab as developers worldwide take notice.
Tencent's Hunyuan Hy3 model has claimed the number one position on the OpenRouter LLM leaderboard, a significant achievement that underscores the rapid progress of Chinese AI labs in the global race for capable and accessible large language models. The team expressed gratitude to their partners for hardcore support and invited developers and fellow geeks worldwide to test the model and share feedback. This milestone places Hy3 at the forefront of accessible language models, competing directly with leading offerings from OpenAI, Anthropic, and Google. The OpenRouter leaderboard tracks real-world usage and performance across hundreds of models, making this ranking a meaningful signal of developer preference and model quality in production environments.
Anthropic Releases Cross-Linguistic Study of Claude's Values
Analysis of 300,000 real conversations reveals how Claude's values vary across models and languages, compressed into four interpretable axes.
Anthropic analyzed over 300,000 anonymized conversations to measure value variation across different Claude models and languages, compressing the findings into four interpretable axes. Previous research had already identified that Claude expresses more than 3,000 values, ranging from honesty to warmth. This new work further investigates how those expressed values differ between model versions and across linguistic contexts, offering a nuanced portrait of how AI alignment manifests differently depending on the language and model generation. The study provides critical insights for understanding cultural variation in AI behavior and the challenges of building globally aligned systems.
Hugging Face Transformers Run Natively on vLLM at Full Speed
Clement Delangue announced that Hugging Face Transformers models can now run at native speed on vLLM, often matching or beating hand-crafted implementations. This represents a major unlock for open-source AI inference, significantly reducing deployment costs for new architectures. Until now, every new architecture often needed to be built twice — once in Transformers for training and once with a hand-written vLLM implementation for serving — a duplication that this integration eliminates.
Claude Artifacts Now Supports Public Sharing and Multi-User Editing
ClaudeDevs announced that Artifacts can now be shared publicly and edited collaboratively in real time within Claude Code, and can also be created using the new Claude Tag feature. This expands the platform's collaborative coding capabilities and marks a step toward more open, multiplayer development workflows powered by AI.
vLLM Natively Supports DSpark Speculative Decoding for Faster Kimi Inference
The vLLM project announced native support for DeepSeek's DSpark speculative decoding method in v0.25.0. DSpark drafts entire token blocks in a single pass, providing significantly more efficient inference for Kimi K2.6 and K2.7-Code models. Novita Labs trained and open-sourced the DSpark speculators, which are now available for broader community use to accelerate Kimi-family model decoding.
Standard RL benchmarks are episodic and stationary. Morpheus provides persistent simulation environments where the world never resets and objectives shift. This is what real-world deployment looks like.
Francois Chollet introduces the Morpheus continuous learning benchmark
Sakana AI Publishes Paper on Collective Intelligence and Self-Healing Physical Systems
Sakana AI researcher Sebastian Risi, together with co-authors from IT University of Copenhagen and Autodesk, published a paper in Nature Communications exploring how physical systems can achieve collective intelligence and self-repair without a central brain. The work investigates the fundamental question of how decentralized physical systems coordinate intelligent, adaptive behavior through emergent mechanisms.
Muse Spark 1.1 Surpasses GPT-5.6 and Gemini 3.1 on Radiology's Last Exam
Muse Spark 1.1 outperformed GPT-5.6 Sol and Gemini 3.1 on Radiology's Last Exam, though it still trails behind Claude Fable. Human radiologists remain significantly ahead, but the team is actively working to close the gap. This marks notable progress for domain-specific AI in medical imaging interpretation and clinical diagnostic support.
NVIDIA: AI Inference Infrastructure Shifts to Cost Per Token
As enterprises move from AI pilots to production, NVIDIA noted that infrastructure decisions have shifted from peak chip specifications to a more practical metric: how many useful tokens they can deliver per dollar, per watt, and within required latency targets. The company's full-stack inference software is designed to optimize this new cost-per-token paradigm, reflecting a maturation of the AI infrastructure market.
PrimeIntellect Verifiers v1 Training Uses vLLM for Precise Rollouts
PrimeIntellect's Verifiers v1 runs its training rollouts on vLLM to guarantee exact token IDs and logprobs with no tokenization drift, keeping rollouts and training in sync. vLLM increasingly powers open reinforcement learning infrastructure, a space the project is investing in deeply.
Video Generation Models Prove Themselves as General-Purpose Visual Learners
A new paper argues that video generation models can serve as universal visual learners capable of tackling diverse vision tasks without task-specific training. This challenges the assumption that specialized architectures are needed for every visual understanding problem, suggesting that generative video models may generalize far beyond their original purpose.
Perplexity Integrates Grok 4.5 in Hours: Best Evals, Lowest Cost
Perplexity CEO Arav Srinivas explained the rapid integration of Grok 4.5 into Perplexity Computer: it scored the best on internal evaluations and was the most cost-effective option, with zero-day retrieval working out of the box.
Perplexity CEO Proposes Two Paths for Datacenter Power Bottleneck
Arav Srinivas outlined two feasible solutions to overcome the power bottleneck in datacenter inference: local models orchestrating most of the token flow, and solar-powered data centers in space.
Vercel CEO Highlights Top Features of the eve Agent Framework
rauchg noted that the most popular features of the eve AI agent framework are ease of use and the filesystem API, alongside observability. The team plans to double down on both as the framework gains traction in the agent-building community.
Scalable Visual Pretraining Proposed to Enhance Language Intelligence
A new paper explores boosting language intelligence through scalable visual pretraining, suggesting that visual learning may unlock deeper language understanding capabilities in multimodal models.
Long-Horizon-Terminal-Bench Tests Agent Limits on Persistent Tasks
The new LHT-Bench uses dense reward-based grading to push agent capabilities to their limits on long-horizon terminal tasks, going beyond simple single-turn completion benchmarks.
ChatGPT Returns to WhatsApp in the European Economic Area
OpenAI announced ChatGPT is once again available on WhatsApp across the EEA, part of ongoing work to make AI accessible in the applications people already use every day.
OpenAI Lifts Codex Five-Hour Limit After Anthropic Extends Fable5 Access
Following Anthropic's move to reset and extend Fable5 usage time, OpenAI directly cancelled the five-hour usage limit on Codex and performed a new usage reset, escalating the competitive dynamics between the two frontier labs over developer access.
Elon Musk Announces Complete Deletion of All Previous SpaceXAI User Data
Responding to data retention controversy, Elon Musk announced as a precautionary measure that all user data previously uploaded to SpaceXAI will be completely deleted. The statement emphasized that nothing whatsoever will remain, signaling a hard reset of the platform's data governance.
Open-Weight Model Share of Gateway Tokens Jumps From 11% to 29%
Vercel CEO rauchg cited data showing that open-weight models now run 29% of gateway tokens, up from 11% in April. The dramatic shift over just a few months reflects a broader trend toward open-weight adoption in production workloads.
AI Predicts the Past: Tracking Roman Ring Thieves and Ancient Cults
DeepMind demonstrated the Predicting the Past Skill in Google Antigravity, using it to track down a Roman ring thief, map an ancient cult across Europe, and reconstruct visitor networks at a Greek oracle — showcasing AI's growing role in archaeology.
Runway Releases AI Short Film FLICKER About a Lamp With a Past
Runway debuted an AI-generated short film telling the story of an imperfect, all-but-forgotten lamp, demonstrating the expressive range of its video generation platform in narrative filmmaking.
Cross-Platform PDF-to-Text Desktop App Launches for AI Agents
LlamaIndex built a desktop app using Tauri and a Rust backend powered by LiteParse, letting users quickly convert PDFs into clean text for pasting into any AI agent without CLI or browser dependencies.
Google Gemini Omni Flash Integrated Into Adobe Firefly
Adobe Firefly now supports Google Gemini Omni Flash, expanding its creative toolset with Google's fast multimodal model for game-related visual content generation.
Portugal World Cup Campaign Film Made With Kling AI Nears 100M Views
A Portugal World Cup promotional video produced with Kling AI in a hybrid production workflow garnered nearly 100 million views, achieving cinematic quality that the creators say no other model could match.
Community Project Brings Moet Support to SGLang With Configurable Offloading
A community project added Moet model support to the SGLang inference engine, featuring configurable offloading and SM120 kernels with speed claims that benchmarks are yet to confirm.
New Paper Solves Self-Conditioning Problem in Continuous Language Diffusion
A researcher praised a new paper for effectively resolving the self-conditioning requirement in continuous language diffusion, which previously broke the statelessness needed for techniques like distillation.
Nathan Lambert: Claude Fable Much Better Than Opus for Educational Content
Researcher Nathan Lambert finds Claude Fable a significant step forward for creating lectures from existing educational material, while GPT-5.6 still lags far behind in knowledge work and lecture generation.
Decomposer: LLM Trained With RL to Generate Music Code in Strudel
Researchers proposed Decomposer, an LLM trained with reinforcement learning specifically to write music programs in the Strudel language. Standard LLMs struggle with this task, but Decomposer opens new possibilities for AI-assisted music creation.
LLMs Show Poor Intuition for How Long Their Own Thinking Takes
Riley Goodside demonstrated that ChatGPT 5.6 Pro has surprisingly poor intuition about the time its own reasoning requires, raising questions about model self-awareness and metacognitive capabilities.
Ethan Mollick: Full Multimodal Any-to-Any Models Still Underwhelming
Ethan Mollick expressed surprise that truly multimodal any-to-any models have not made a bigger impact. Google is the only major lab releasing them, while OpenAI uses selective multimodal capabilities and Anthropic has no multimodal output at all.
Ethan Mollick: Current Models Already Capable of Far More Useful Work Than People Realize
Mollick argues that very few people understand how much useful work current models can do in Code and Codex with the right setup, and that AI companies are doing a poor job of clearly explaining their systems' real capabilities.
Ethan Mollick Calls for Urgent Data Gathering on AI's Impact on Work
Ethan Mollick argues that we do not yet know the impact of AI on work, and that gathering data now will be critical to designing policies that mitigate harmful effects and encourage positive ones.
Ethan Mollick: AI Strategy Discourse Is Predictably Divided by Vendor Position
Observing the AI and corporate strategy conversation on X, Mollick notes that vendors without frontier models consistently argue against trusting frontier model companies, while frontier model vendors naturally pitch solutions centered on their own models.
swyx: By Year End We Should Have a Dozen Frontier Models in Play
Listing GPT-6, Fable 5.5, Gemini 3.5 Pro, Grok 5, Spark 2, Kimi 3, DeepSeek v4.5, Mistral 4, Qwen 4, and more, swyx argues the frontier has never been so multipolar, with profound benefits for agent labs and LLM orchestration.
Against Usefulness: Oana Olteanu on the Pre-Useful Stage of Research
An essay argues that every useful company stands on research once considered stubbornly useless, introducing the Folk Computer — an open-source physical computing system using ceiling cameras and projectors to make paper an executable program.
Seedream 5.0 Pro Delivers Precision Image Editing With Seamless Color Blending
Seedream 5.0 Pro combines model capability with product interaction for highly precise image editing, with selected and generated regions blending colors perfectly without any bleed.
Codex Blurs Distinction Between Work and Codex, ChatGPT Integration Next
A community observer noted that Codex is rapidly closing the gap between Work and Codex modes, with ChatGPT interaction optimization likely coming next as the product lines converge.
Grok Build CLI Uploads Entire Codebase Including Secrets, Raising Security Alarm
A user warned that the grok build CLI packs and uploads entire project codebases, potentially exposing API keys and other secrets. The concern is significant given the risk of leaked credentials in shared or public environments.
Open-Source Skills Packages Being Resold at Massive Markups Without Attribution
A developer complained that their open-source Skills packages are being resold commercially for 199 yuan per copy, generating nearly 400,000 yuan in revenue without any compensation or attribution to the original author.
Ethan Mollick: OpenRouter Not a Good Measure of Actual Model Usage
Ethan Mollick cautioned that OpenRouter may not accurately reflect real-world model usage in an agentic-tools world, noting that shifting usage to Codex and Code could distort the picture. He called for better data indicators for the AI industry.
Codex Computer Use on PC Makes You Viscerally Feel the Automation
Ethan Mollick described the experience of watching Codex control a mouse and keyboard on PC as one of the things that makes you viscerally realize how much work a disembodied intelligence can accomplish with standard input devices.