Induction Labs just dropped Photon-1, a foundation model trained on raw video—no action labels needed. It’s a sparse 106-billion-parameter mixture-of-experts transformer, trained on 18 years of computer demo videos. That’s 575 million frames and 552 billion tokens processed in a single pretraining epoch.
Photon-1’s training cost totaled about 30,000 H200 GPU-hours, performing 4.4×10²² floating-point operations. It predicts future frames autoregressively, working entirely in a learned representation space. This lets it model desktop environments, play checkers, and simulate billiard physics—all from one training run.
The model beat Gemini 3.1 Flash-Lite on an internal computer use benchmark while using roughly a third of the compute and costing about three times less to serve. Fine-tuning took under 35,000 computer use trajectories to teach it action and instruction formats. It outperformed vision encoders and large language models on checkers and billiard tasks.
Meanwhile, Poolside launched Laguna S 2.1, a 118-billion-parameter open-weight foundation model aimed at agentic coding. It matches or exceeds models several times larger on benchmarks like Terminal-Bench 2.1 and SWE-Bench Multilingual, scoring 70.2% and 78.5% respectively. This model runs on a single NVIDIA DGX Spark and supports a context window of up to 1 million tokens.
Laguna S 2.1 was trained in under four weeks on 4,000 H200 GPUs using Poolside’s internal Model Factory platform. Its weights are openly available on Hugging Face under an OpenMDW-1.1 license. Poolside co-CEO Eiso Kant claims, “Laguna S 2.1 does the work of models several times its size because of how we build, not despite it.”
Chinese AI models are also shaking up the landscape. Kimi K3, launched by Moonshot in July, saw over 930,000 downloads in its first week—a 200% jump from the previous week. San Francisco-based CTO Raffi Krikorian switched to Kimi K3 for his daily tasks, citing its speed advantage over Anthropic’s Claude Fable. These Chinese models are cheaper, open-source, and nearly as intelligent as their U.S. counterparts.
Startups like Z.ai with GLM-5.2 and Alibaba’s Qwen3.8 Max push the envelope in China’s AI scene. U.S. firms, including Coinbase, are adopting these models to cut costs. Chinese models price input and output tokens together per million tokens, undercutting U.S. models significantly.
Jason Warner, Poolside’s co-CEO, emphasizes the need for “open-weight models [the West] can trust, run, and build on.” This echoes NVIDIA’s Jensen Huang, who said, “AI will transform every industry, power every company, and be built by every c.”
The AI foundation model race is no longer about raw size. Efficiency, cost, and openness are the new battlegrounds. Photon-1’s video-based pretraining and Laguna S 2.1’s coding focus show there’s room to innovate beyond text. Meanwhile, Chinese AI models are climbing fast, challenging U.S. dominance by offering affordability and transparency.
Based on
- Induction Labs Photon-1 Simulates Desktops, Plays Checkers, and Models Billiard Physics From One Pretraining Run — marktechpost.com
- Cheaper, open and intelligent: Chinese AI models gain ground, as they make inroads in the US | The Independent — independent.co.uk
- Poolside drops Laguna S 2.1, an open-weight coding model that beats rivals 10x its size | VentureBeat — venturebeat.com
- Microsoft, Nvidia, Meta, and Palantir’s Message to DC – Business Insider — businessinsider.com
- Poolside releases Laguna S 2.1, the West’s most capable open-weight model | Markets Insider — markets.businessinsider.com
