xAI Ships Grok 4.7, OpenAI Forms an Independent Math Advisory Group After Its Navier–Stokes Result, Xiaomi Open-Sources MiMo-V2.6 and Alibaba Releases Qwen-Image-2.1

Posted on by

This brief covers the trailing ~72 hours (September 19–22, 2026). Every item below was confirmed on the originating organization’s own page, with a published date inside the window. The window was a model-release weekend: xAI shipped Grok 4.7 as its new flagship for coding and knowledge work, Xiaomi open-sourced the trillion-parameter, omnimodal MiMo-V2.6 series, and Alibaba’s Qwen team released Qwen-Image-2.1, a compact unified image generator and editor with native transparency. On the governance side, OpenAI responded to an open letter from mathematicians by standing up an independent Advisory Group on Mathematics and AI to review and communicate results from the internal model that resolved Navier–Stokes earlier this month. Google opened pre-orders for Googlebook, a laptop built around on-device Gemini, and OpenAI expanded OpenAI Academy with role-based learning paths.

xAI releases Grok 4.7, a larger base model with a new safeguard stack, at the same $2/$6 price as Grok 4.6

SpaceXAI · September 21, 2026

Grok 4.7 uses a new, larger base model than Grok 4.6 and was trained with a longer reinforcement-learning run weighted toward tasks that take many hours to complete, with explicit training to understand the Grok Bot harness. xAI reports 46.3% on CursorBench 4.0 (vs. 40.4% for 4.6), 71.0% on DeepSWE v1.1 at high effort, 38.0% on Terminal-Bench 4.0 (up from 20.3%), and a GDPval Elo of 1,695, placing it between Grok 4.6 and Fable 5.1. The company says the model was built with an entirely new safeguard stack, topping LatchBio’s biosafety benchmark at 62.4% and allowing only 3.3% of risky dual-use prompts through on its HackerBench v0.3, while select cybersecurity partners get invite-only access to its red-team capabilities. It is available today in Cursor, Grok Build, and the Grok API at $2 per million input tokens and $6 per million output tokens, with a fast variant at twice the output speed for twice the price.

“It works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date.” — SpaceXAI

Source: Introducing Grok 4.7

OpenAI stands up an independent Advisory Group on Mathematics and AI after its internal model resolves 100+ open problems

OpenAI · September 21, 2026

OpenAI disclosed that the internal model it began training on August 28, the same system behind its Navier–Stokes result, has now resolved more than 100 long-standing open problems across most areas of mathematics, at a pace that surprised the company’s own mathematicians. Citing the open letter “A Severe Misalignment of AI in Mathematics,” in which mathematicians objected to solving open problems as a benchmark for new AI systems, OpenAI said it is working with an independent advisory group hosted at the Institute for Advanced Study to assess the significance of emerging results, coordinate their dissemination, and advise on academic standards. Members are unpaid, may publish unsolicited advice, and can change the group’s membership; initial members include Timothy Gowers, Martin Hairer, Edward Witten, Ravi Vakil, Camillo De Lellis, Melanie Matchett Wood, Ulrike Tillmann, Nikhil Srivastava, and François Charles. OpenAI notes the group will not advise on how to pace its internal progress on mathematics.

“The group will operate independently from OpenAI. The group will have the freedom to offer advice we have not requested, comment on OpenAI’s impact on mathematics, and make its advice public.” — OpenAI

Source: Advisory Group on Mathematics and Artificial Intelligence

Xiaomi open-sources MiMo-V2.6: a 1.02T-parameter omnimodal MoE with 1M context, trained in one mixed RL run

Xiaomi MiMo · September 21, 2026

Xiaomi released the MiMo-V2.6 series under an MIT license, led by MiMo-V2.6-Pro-RL, a sparse mixture-of-experts model with 1.02 trillion total and 42 billion active parameters, a 1M-token context window, and native text, image, video, and audio input. The technical approach centers on “You Only RL Once,” a single mixed reinforcement-learning run spanning coding, general agents, vision, and cybersecurity, plus a groupwise agentic grader that ranks passing rollouts against each other to push toward shorter, cheaper solutions. Xiaomi’s own evaluation table puts Pro at 71.9% on DeepSWE v1.1, 76.9% on Toolathlon-Verified, 82.0% on OSWorld-Verified, and 94.0% on CyberGym, generally within a few points of Claude Opus 5 and GPT-5.6 Sol on agentic benchmarks. A smaller Flash model (309B parameters) and an UltraSpeed variant of Pro are served through Xiaomi’s API at $0.435/$0.87 and $0.14/$0.28 per million input/output tokens respectively.

“One mixed RL run across coding, general agents, visual, and cybersecurity — not separate per-domain runs.” — Xiaomi MiMo Team

Source: MiMo-V2.6 (see also the MiMo-V2.6-Pro-RL model card)

Alibaba’s Qwen team releases Qwen-Image-2.1, a 7B unified generator and editor with native transparency

Qwen (Alibaba) · September 20, 2026

Qwen-Image-2.1 unifies text-to-image generation and image editing in a single model with just 7 billion parameters in its visual generation component (a 32-layer single-stream DiT), paired with a Qwen3-VL 8B text encoder and a 64-channel RGBA VAE. It natively generates and edits transparent images, accepts up to 10 reference images for multi-subject composition, supports circle, paint, and mask annotations to target local edits, and renders natively at 2K resolution. A mixed-granularity attention design lets the model encode text and condition images once and reuse the prefix KV cache across all denoising steps. Weights are on Hugging Face and ModelScope under the Qwen Research License, with day-zero support in Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V, and two fine-tuned Qwen3.5-VL 9B prompt-rewriting models released alongside.

“We are excited to open-source Qwen-Image-2.1, a unified text-to-image generation and image editing model in the Qwen family.” — Qwen team

Source: Qwen-Image-2.1 on GitHub (blog: qwen.ai)

Google opens Googlebook pre-orders: a $899 laptop built around on-device Gemini, Magic Pointer, and Antigravity

Google · September 21, 2026

Google detailed the intelligence layer of Googlebook, its new Android-and-ChromeOS-based laptop, which brings Gemini directly onto the device. Magic Pointer summons Gemini with a cursor wiggle to act on whatever is on screen (scheduling a training plan into Calendar, checking whether a hovered email is spam, combining selected images); Rambler turns spoken stream-of-consciousness into structured, multilingual notes; and Create My Widget builds custom widgets from a description. Every Googlebook ships with Google Antigravity and a full Linux terminal for agentic coding tools, and Gemini Spark can keep processing tasks after the lid is closed. Pre-orders start at $899 and include 12 months of Google AI Pro, with devices arriving October 4 in the U.S. and October 5 in Canada, the U.K., Ireland, France, Germany, and Australia.

“Developers also have access to a full Linux terminal environment to run tools like Claude Code or Antigravity CLI and do serious agentic coding right on your Googlebook.” — Alexander Kuscher, Google

Source: Googlebook’s built-in intelligence reinvents the way you use your laptop

OpenAI expands OpenAI Academy with role-based learning paths and course badges

OpenAI · September 21, 2026

OpenAI added new course pathways to OpenAI Academy for developers (Build with AI, eight courses covering Codex and the API, evaluations, agents, and production operation), leaders (an AI Leadership course on strategy, ownership, and roadmaps), and educators and college students (AI for Educators and AI for College Students), joining the existing Apply AI at Work pathway for knowledge workers. Each course ends with an assessment, and learners who pass earn an OpenAI Academy course badge. The company positions the expansion as part of deployment, encouraging organizations to combine pathways for onboarding, technical teams, and executive programs.

“At OpenAI, we treat learning as part of deployment.” — OpenAI

Source: Expanding OpenAI Academy with new learning paths

Still developing

OpenAI publishes an Australian Youth Safety Blueprint (September 18, 2026). Just before this window opened, OpenAI released a six-pillar roadmap for protecting young Australians using AI, spanning AI literacy, age-appropriate safeguards, privacy-protective age assurance, connections to crisis support, and parental controls, and noted that ChatGPT for Teens began rolling out in Australia in August as the default experience for users identified as 13 to 17. Source: Introducing the Australian Youth Safety Blueprint

This brief covers the trailing ~72 hours (September 19–22, 2026).

Primary sources:

Leave a Reply