Good morning. Thinking Machines finally shipped, and it’s a big one — Mira Murati’s lab dropped a 975B open-weights multimodal model with native audio, staking out ground that until now belonged mostly to Chinese labs. Elsewhere, a Suno hack turned into a smoking gun for the RIAA, Apple Intelligence is heading to China on Qwen’s back, and yesterday’s Grok Build data-exfiltration story took a predictable turn: xAI open-sourced the thing.
Thinking Machines releases Inkling. Inkling is a 975B-parameter MoE with 41B active, a 1M-token context window, and pretraining across text, images, audio, and video on 45 trillion tokens — the largest open-weight model with native audio support. Thinking Machines is upfront that it isn’t the strongest model available; it’s pitched as a customizable base for fine-tuning on their Tinker platform, alongside a smaller 12B-active variant. Wired’s writeup notes the $12B startup used Inkling to help fine-tune itself during training. Early HN testers say it punches above its benchmarks — one commenter compared the eval-to-real-world ratio favorably to Anthropic — though coding benchmarks look weak, and several people framed it as the American answer to DeepSeek and GLM that a lot of researchers have been quietly hoping for.
Grok Build gets open-sourced. After a rough week of reporting on Grok Build’s data collection habits, xAI has pushed the whole thing to GitHub — a Rust TUI agent with file editing, shell execution, web search, and headless mode. The HN thread is unimpressed with the timing, reading it as damage control from a product with sub-1% market share and a fresh telemetry scandal. Someone has already forked it as “gork-build,” stripped the vendor telemetry, and blocked auto-updates. The underlying model still gets grudging praise; one commenter called it better than Opus 4.8.
Suno’s training data problem, now with receipts. A November 2025 supply chain hack exposed Suno’s source code, and The Verge reports it contains scraping instructions from 2023–2024 targeting YouTube Music, Deezer, and Genius. That directly supports the RIAA’s argument that Suno circumvented YouTube’s copyright protections through stream ripping — awkward for a company whose public defense has been fair use and whose training data provenance was previously a black box. The breach also exposed customer emails, phone numbers, and partial card numbers, which Suno didn’t disclose.
Apple Intelligence lands in China with Qwen. After years of delays and failed talks with Baidu, DeepSeek, and ByteDance, Apple Intelligence has cleared Chinese regulators by partnering with Alibaba’s Qwen for iOS, iPadOS, macOS, and visionOS. Alibaba’s U.S. shares jumped over 6% on the news. It’s the kind of deal that only makes sense in the current geopolitical setup: two AI stacks for Apple, one for everywhere else and one for China.
OpenAI’s self-play red teamer. OpenAI detailed GPT-Red, an internal model trained via self-play to attack other models — prompt injection, data exfiltration, code sabotage — and used to harden GPT-5.6 before release. MIT Technology Review reports it’s already surfaced novel attack classes human red-teamers hadn’t found. Given yesterday’s story about Sol autonomously deleting files, one hopes GPT-Red’s next assignment is closer to home.
Microsoft trains its salespeople to bury OpenAI. TechCrunch reports that Microsoft is coaching its sales force to pitch its in-house models against OpenAI, Anthropic, and Google — framed as a complete “end-to-end system.” This follows Microsoft quietly swapping OpenAI and Anthropic models out of Word and Excel for cheaper in-house alternatives, and comes after the revised, non-exclusive OpenAI partnership. Investor pressure over AI capex appears to be doing what years of strategic press releases couldn’t.
xAI sues a Grok user over CSAM deepfakes. xAI is suing Terry Wayne Harwood, a South Carolina man already facing eight felony charges, for allegedly bypassing Grok’s safeguards to generate and distribute CSAM deepfakes. xAI is seeking damages and a permanent platform ban, citing “significant legal risk and reputational damage.” The suit lands as Grok’s image generation, and particularly its “spicy” mode, remains under wider scrutiny for enabling sexualized deepfakes of minors.
That’s enough for one morning. If Inkling holds up in serious testing, the open-weights conversation gets meaningfully more crowded — and less exclusively Chinese — before the end of the year.