Inception Labs' Mercury 3 Hits 981 Tokens per Second on Reasoning Tasks
Inception's Mercury 3 Preview became the first diffusion LLM scored on ARC-AGI, averaging 981 tokens/sec and peaking at 1,743.
20 stories
Inception's Mercury 3 Preview became the first diffusion LLM scored on ARC-AGI, averaging 981 tokens/sec and peaking at 1,743.
GlobalFoundries becomes the first semiconductor fab to pilot the Atlas humanoid, while scaling up a Spot fleet that already hit a 99% mission success rate.
Google Colab now exposes a REST API and a hosted MCP endpoint, letting agents provision GPU and TPU runtimes and run Python headlessly.
A 4.5M parameter English TTS model trained only on synthetic speech runs 70x faster than real time on CPU and cost under $10 to build.
A new theory explains why x-prediction wins in high dimensions, and k-Diff learns the ideal target automatically with one extra parameter.
Infomaniak repackaged Google's TranslateGemma 4B so it runs cleanly on vLLM and SGLang, adding a delimiter-based language code format for serving.
Google Cloud AI Research has open-sourced a multi-agent system that turns experiment logs and raw ideas into submission-ready LaTeX manuscripts with real citations.
A new measurement-based quantum computing approach shows that just one extra trainable parameter is enough to beat purely unitary generative models.
Kaggle and HeyGen turn the Code2Video benchmark into a public head-to-head, where humans vote on which model wrote better motion graphics code.
A self-hosted, Apache-2.0 alternative to ChatGPT Enterprise that puts shared agents, RAG, a real browser and 4,700+ MCP servers behind your own firewall.
A Google framework treats the agent runtime as mutable code, showing modular self-edits with validation gates beat full rewrites on SWE-bench Lite at a fraction of the token cost.
Princeton researchers show that sequence-level on-policy distillation can fixate on late tokens and nudge students toward refusals the teacher never made.
Researchers from MIT and NVIDIA hand quantization recipe design to a coding agent, using a minutes-long KL proxy to steer the search.
NVIDIA has released an open 120B multimodal model, a 50,000-question chemistry benchmark, and reinforcement learning recipes to turn Nemotron into domain-specific science agents.
Alibaba's Qwen4-Max is a 2.4T-parameter agent model that cuts per-task costs by about 60%. It went live on You.com's Inference Router on day one, but its weights stay closed.
LM Studio's Bionic agents can now run on remote laptops, dev boxes, and cloud VMs, controlled from any device over an encrypted Tailscale mesh.
Claude Code Desktop now drives Apple's iPhone Duo simulator, folding and unfolding the device to test how your app adapts to the foldable screen.
Databricks and Turing built a benchmark showing Genie with its Ontology layer beats two frontier agents 91% to 45-58% on enterprise data tasks.
OpenAI brings computer-use agents to clinical software with CVS, Cigna, Abridge and Ambience, plus 20+ new plugins and improved clinical search.
Hark Labs is bundling a 5% cashback virtual card with its AI agent, capped at 10,000 invite-only members, turning its assistant into a payment layer.