Summary
Matthew Berman rounds up a busy week of open-weights model releases, led by Moonshot’s release of the Kimi K3.1 Pro weights. The roughly 3 trillion parameter model ships in 96 parts of about 17 GB each and needs around 1.2 TB of VRAM, so most people will rely on inference providers until quantized versions appear. Reported benchmark results place it ahead of Astra on Terminal Bench, first on Agents Last Exam, third on OSWorld, and second on GDPval behind Opus 5.5.
The video also covers North Ka, a family of open computer-use models in 397B, 35B, and 9B sizes with its own open-source harness from Beo AI. The reported OSWorld Verified scores are 88 for the largest model, 85 for the Mini, and 78.9 for the Nano. Other releases include Pruna AI’s image generation and editing model built from MiniMax H3, which is said to run five times faster in eight steps, and MIT’s Hydrogen 1.0, a science-focused model post-trained from GLM 5.3 that scores 38.33% on Frontier Science at about 25 cents per problem.
The roundup closes with Perplexity’s open-weights vision-language embedding models in 9B and 600M sizes, aimed at retrieval over text, images, and PDFs, and a community-tuned 4B Qwen 3.5 coding model that generates Three.js scenes.
📺 Source: Matthew Berman · Published November 05, 2026
🏷️ Format: Roundup
