Company@X — 2026-07-14#

Signal of the Day#

Meta AI demonstrated a major leap in frontier model reasoning capabilities, achieving a perfect 30/30 score on the Asian Physics Olympiad’s theoretical exam to tie with the top three human contestants. This confirms that advanced multimodal models are crossing new thresholds in rigorous, multi-step scientific reasoning.

Key Announcements#

Google · Source Google is bringing its latest Nano Banana model directly into Search via AI Overviews, allowing users to generate custom images on the results page. The company is also rolling out a dynamic, real-time gallery revamp for Google Images on desktop. This marks a significant integration of generative visual capabilities into core search workflows rather than siloed applications.

Chai Discovery · Source Chai Discovery secured a $400M Series C at a $3.8B valuation, backed by Index Ventures, Sequoia, and Kleiner Perkins. The massive funding round signals high capital conviction that AI-driven drug discovery is transitioning from theoretical promise to scalable, engineered deployment.

Tencent / Hy3 · Source A new 298B-parameter Hy3 model release features aggressive 1-bit and 2-bit FPX quants that allow the massive architecture to run on a single 128GB GPU or AMD Strix Halo hardware. This pushes the boundary of local inference, enabling flagship-scale intelligence on radically constrained hardware.

Singularity · Source Hard-tech defense startup Singularity unstealthed with an $80M Series A at a $400M valuation. Founded by former SpaceX and Anduril engineers, the company is building low-cost air defense interceptors engineered explicitly for automotive-scale mass production rather than bespoke defense manufacturing.

PrismML · Source PrismML open-sourced Bonsai 27B, positioning it as the first 27B-class model optimized to run locally on smartphones. By utilizing extreme quantization down to 1.125 effective bits per weight, the 3.9GB footprint brings multi-step reasoning and agentic workflows directly to mobile edge devices.

Anthropic · Source Anthropic launched Claude for Teachers, offering verified K-12 educators in the US free access to premium model capabilities alongside evidence-based curricula mapped to academic standards. Separately, the company committed $10 million CAD to fund new AI research in partnership with Canadian institutions.

Also Noted#

• Open_MOSS (Source): Open-sourced MOSS-VL-Realtime, an 11B vision-language model designed to continuously ingest video streams and instantly revise or interrupt its responses as scenes change.
• Prediction Markets (Source): Launched GPU compute forward curves for Nvidia B200, H200, and A100 chips, providing critical market-implied pricing infrastructure to treat compute as a tradable commodity.
• OpenClaw (Source): Released version 2026.7.1 with major iOS and Android app overhauls, alongside newly added support for GPT-5.6 and Muse Spark 1.1.
• Ideogram (Source): Open-sourced its V4 models (Fast and Instant variants), making them immediately available for download on HuggingFace and inference on fal.
• LeRobot (Source): HuggingFace’s robotics library added real-time, end-to-end depth perception support, seamlessly encoding 12-bit depth video alongside RGB streams for robotic training.