Podcast Episode
Open Season: Kimi K3's Price War, NVIDIA's 25x Video Leap, and an AI-Cracked Maths Mystery
July 23, 2026
0:00
8:32
A 'quiet' news day still delivered a flood of open models and faster tools. We dig into Moonshot's Kimi K3 undercutting Western labs, Upstage's Solar Open2, NVIDIA Cosmos 3's 25x faster video, Microsoft's screenshot-driven Fara1.5 agent, the Gigatoken speedup, a US Department of Energy science model, an AI-assisted counterexample to a 30-year-old maths conjecture, and Cursor's cost-cutting model router.
The Open-Weight Price War: Moonshot's Kimi K3
The day's dominant model story came from Chinese lab Moonshot, whose open-weight Kimi K3 is being downloaded and run locally at scale. One independent analysis put the top-tier K3 roughly level with a leading Western closed model on hard software-engineering tests at about 55% of the price, and one coding tool reported K3 leaping from 0% to 16% of its usage in just three days. The controversy: White House science advisor Michael Kratsios alleged Moonshot built K3 by distilling one of Anthropic's Claude models (Fable). Many technical observers pushed back on feasibility, noting the Claude model was available for only days before K3 shipped.Korea's Upstage Ships Solar Open2
Upstage released Solar Open2, a 250B-parameter open model, extending a striking 2026 trend: the open shelf is now crowded, and much of the momentum is coming from Asian labs. More open competition generally means lower prices and more choice for businesses building on top.NVIDIA Cosmos 3: Video Generation 25x Faster
NVIDIA announced Cosmos 3 Super, generating images and video up to 25x faster while ranking near the top of open-weight leaderboards, plus Cosmos 3 Edge for physics-aware video understanding on smaller devices. The edge angle matters for robots, vehicles, and cameras that must interpret the world in real time on-device.Microsoft Fara1.5 Browses Like a Human
Microsoft's Fara1.5 is a computer-use agent that drives a web browser purely from screenshots, no DOM or accessibility tree, emitting clicks, typing, and scrolls. Fine-tuned from a Qwen base, it points toward assistants that book appointments or fill forms, though Microsoft candidly flags prompt-injection risk from malicious page content.Gigatoken: The Unsexy 100x Speedup
Gigatoken is an open-source tokenizer claiming roughly 100x faster throughput than the standard tool. It won't speed up a single chat, but for pipelines chewing through millions of documents, it turns overnight jobs into coffee-break ones.Genesis-Science-1: AI Meets the US Department of Energy
American firm Arcee announced a partnership with the US Department of Energy to build Genesis-Science-1, a trillion-parameter-class American open model with a governed research harness aimed at reproducible scientific computing, a serious institutional bet on open AI for science.AI Cracks a 30-Year-Old Maths Conjecture
Researcher Dmitry Rybin used a frontier model to help find a counterexample to the Dinitz-Garg-Goemans conjecture, an open graph-theory problem roughly 30 years old. The model generated candidates; a human expert verified the real result, a template now producing a firehose of plausible research artefacts.Cursor Router and the Cost of Intelligence
Cursor launched Cursor Router, which routes each task to the cheapest capable model and reserves premium models for the hard jobs, claiming frontier-quality results at around 60% lower cost. Smart routing is fast becoming table stakes for high-volume coding and agent workloads.Published July 23, 2026 at 6:50am