Sam Altman said on the Sources podcast that OpenAI will definitely make a humanoid robot, ending months of speculation about whether the company's hardware push included physical bodies.
Trained directly on real-time satellite and weather-station observations, WeatherNext 3 produces hourly forecasts at 5 km resolution and cuts precipitation error by up to 50%. It is already live across Google's flagship products.
xAI opened Grok Bot to enterprise customers, letting whole organizations — including people without an existing seat — use persistent agents that live beyond a single session. Grok and Cursor Enterprise customers get free usage for two weeks.
Randomly evicting KV cache entries within each head matches the strongest prior evictor across four models and six reasoning tasks while serving 32-43% higher throughput in vLLM, suggesting eviction selection signals add almost nothing.
A 6B diffusion transformer trained from scratch with a frozen diffusion-VLM module sets a new open-source state of the art on Qwen-Image-Bench and distills to 2-4 step Turbo inference.
Researchers reconstruct 37,300 executable terminal environments from recorded agent trajectories; SFT of Qwen3.5-27B improves Terminal-Bench 2.1 by 11.9 points and EvoCode-Bench v2 by 13.8 points.
Restart-controlled truncation probes show that apparent 'breakthrough' moments and early-legible fates in LLM reasoning traces are confounded by test-time budget and problem difficulty, not by path properties.
A Google DeepMind paper lets models declare global, focus or local attention modes in chain-of-thought so the runtime skips KV cache reads, cutting attended tokens by 52% on Gemma-4-31B and 31% on Qwen-3.6-27B across 15 long-context tasks.
K2 introduced Horizon, six open models released together as one connected family, expanding the lineup beyond the original flagship.
Alibaba's Qwen 3.8 27B dense multimodal model is live on Cerebras' inference platform, targeting agentic coding, tool use and research workloads.
Hermes Desktop now detects GPU, VRAM, CPU and RAM on first launch, then recommends, downloads and configures the best local model and runtime, with NVIDIA-optimized support.
A Rust-based operating system built around AI agents, hosted on GitHub as an open-source alternative to agent-constrained platforms.
At IFA 2026 AMD unveiled the Ryzen AI Max Pro 400 platform and the liquid-cooled Threadripper Halo Station — a 96-core workstation with two MI350P accelerators and up to 576GB HBM3E — claiming it can run models beyond one trillion parameters fully offline. Lenovo and HP showed first OEM systems.
NVIDIA previewed RTX Spark Windows PCs from Lenovo and Acer for October with up to 1.9x faster local inference, and launched PAIR, a Personal AI Router that spreads inference across idle PCs on a home network.
One of the first compact systems on AMD's new professional Ryzen AI Max+ silicon runs a 284-billion-parameter DeepSeek V4 Flash model fully locally.
The developer tool takes text, images, video and 3D assets and renders them as photorealistic, camera-controlled HD video.
A no-code platform for creating agents that recover from failures on their own instead of stopping mid-task.
The open-source audio editor ships a modern Qt interface, a new clip editing model and a revised project file format.
A Google Analytics-style dashboard for model context protocol servers, tracking usage and quality of MCP integrations.
Organizations define the rules and the AI system flags violations automatically in video content.
The AI data center developer is said to be raising $3B co-led by Atreides Management and Valor Equity Partners with Mubadala Capital, serving customers including Meta, Microsoft and OpenAI.
Mira Murati's frontier lab is reportedly raising $1B at a ~$40B valuation with more than $100M in annual revenue, up from a $2B seed at $12B.
The multi-chip inference startup raised $300M led by Andreessen Horowitz, with new backing from Arm Holdings and Microsoft's M12, following an $80M Series A led by Menlo Ventures.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
