Moonshot AI published Kimi-K2-Instruct-0905 on Hugging Face: 1 trillion total parameters with 32.6 billion active per token across 384 experts, a context window extended from 128K to 256K tokens, under the family's modified-MIT license with no revenue gate. The checkpoint targets agentic work — published figures include 55.9 on SWE-bench Multilingual and 44.5 on Terminal-Bench — and Moonshot quotes a hosted API at 60-100 output tokens per second. The weights are downloadable and fine-tunable, and the license permits commercial derivatives.

On the same day, Alibaba announced Qwen3-Max-Preview: also more than 1 trillion parameters, also sparse MoE, text-only and closed, reachable only through Alibaba Cloud and OpenRouter, with self-reported benchmark wins over Claude Opus 4, Kimi K2 and DeepSeek-V3.1. Same scale, opposite distribution. One is a file you can quantize, fine-tune and run on your own hardware; the other is a tiered rate card with no download button.

Warner Bros. Discovery sued Midjourney in Los Angeles federal court, alleging the image and video generator trained on its works and serves subscribers near-identical renderings of Superman, Batman, Wonder Woman, Bugs Bunny, Scooby-Doo and Cartoon Network characters. The complaint seeks disgorgement of profits or statutory damages of up to $150,000 per infringed work, plus an injunction, and attaches side-by-side images of its characters against Midjourney's outputs — down to details like the color of Scooby-Doo's collar.

It follows the June suit filed by Disney and Universal, which shared counsel with Warner Bros., and turns on the same unresolved question: whether training on and reproducing protected characters is fair use. Warner Bros. leans on timing as evidence of intent — Midjourney previously blocked copyrighted-character video generation, then lifted the restriction in August — and the Motion Picture Association has publicly backed the studios.

Anthropic completed a $13 billion Series F led by ICONIQ, with Fidelity Management & Research and Lightspeed Venture Partners co-leading, at a $183 billion post-money valuation — up from $61.5 billion in March. Run-rate revenue went from roughly $1 billion at the start of 2025 to more than $5 billion by August. The company also disclosed that Claude Code, fully launched in May, is already past $500 million in run-rate revenue with usage up more than 10x in three months.

That last number is the one worth keeping. Anthropic is not monetizing chat volume at the top of the funnel; it is monetizing long, tool-calling agent sessions priced per token, and the round is earmarked for compute and enterprise capacity to serve them. A three-month-old CLI is now a half-billion-dollar revenue line, which tells you where the margin sits in the stack.

EPFL, ETH Zurich and the Swiss National Supercomputing Centre released Apertus, Switzerland's first large-scale open multilingual LLM, in 8B and 70B dense sizes under the permissive Apache 2.0 license. It was trained at CSCS on 15 trillion tokens spanning more than 1,000 languages, roughly 40% of the data non-English and including Swiss German and Romansh — languages that mainstream models treat as afterthoughts.

The distinguishing claim is completeness rather than benchmark placement. Architecture, weights, training data and training recipes are all published, along with a peer-reviewed mechanism that the team says prevents verbatim memorization of training data. The models download from Hugging Face, and Swisscom serves them on a Swiss-hosted platform, which makes this a self-hosting option that does not route through a US or Chinese vendor's terms.