The Day in AI

• Microsoft AI finished pre-training MAI-Base-2, a 2.8 trillion parameter base model with 80 billion active parameters, trained on 16,000 GB200 GPUs for 38 days.
• Meta released Muse Spark Open, an 816B-parameter open-weight model with 42B active parameters, which it calls the most capable US open-weight model.
• NVIDIA released Nemotron 3.5 Super VL, an open 120B multimodal model with 12B active parameters, listed on OpenRouter at $0.10 per million input tokens.
• Apple says Apple Intelligence and Alibaba's Qwen have regulatory approval in China, with a beta due in late October.
• Anthropic is giving researchers from the evaluator METR employee-level access to its models and safety practices, TechCrunch reports.

Microsoft says it trained MAI-Base-2 on about 50 trillion tokens, using 16,000 Nvidia GB200 GPUs for 38 days in August and September. It is a mixture-of-experts model, so each token is routed to a few expert subnetworks and only 80 billion of the 2.8 trillion parameters run at a time. It follows MAI-Base-1 from June, with about 36% more unique training tokens. Images are now trained in natively and make up 18% of the mixture.

Microsoft reports that its training recipe is 6.3 to 12.7 times more compute-efficient than MAI-Base-1. Across four evaluations, it puts the gain over DeepSeek-V4-Pro-Base at 4.2 to 12.3 times. Only a paper is out, with no weights, no API and no independent results. The announcement gives no date for the post-trained model.

Meta released Muse Spark Open, three days after its closed Muse Wonder models on 12 October. It has about 816B parameters with 42B active, takes text, images, audio and video, returns text, and has a context window of up to 1 million tokens. Meta trained it for agentic coding and tool use and ships it under a restricted license. On Meta's own numbers it scores 61.8 on JobBench against 61.4 for GLM 5.3, and 89.2 on MCP-Atlas against 84.2. It trails on Terminal-Bench 2.1, at 76.4 against 85.0 for Kimi K3. The model card also reports 39 on the Artificial Analysis Intelligence Index, below GLM 5.3 at 45, so Meta's "most capable" claim holds among US models only.

NVIDIA released Nemotron 3.5 Super VL, an open 120B model with 12B active parameters for long-running enterprise agents. It reads text, images, documents, GUI screenshots and video. It mixes Mamba layers, which process long sequences at a cost that grows linearly with length, with standard Transformer attention layers, and it predicts several tokens per step to speed up generation. NVIDIA says it is 5x faster than the next best open model in its class, but the benchmark behind that figure has not been checked. OpenRouter lists it at $0.10 per million input tokens and $0.40 per million output tokens, plus a free endpoint. NVIDIA also announced two edge models, Nemotron 3.5 Bolt 74B-A9B 24GB and Bolt E4B 4GB.

• Apple says Apple Intelligence will use Alibaba's Qwen and Tongyi Wanxiang models in China, with a wider launch reported for December in iOS 27.3 and no redesigned Siri.
• Mistral AI released Timestral, its first time-series forecasting models in four sizes from 5M to 209M parameters, and says the 209M model leads zero-shot on GIFT-Eval.
• Google launched Universal Cart in the US, which saves products from several retailers across Search AI Mode and the Gemini app.
• Amazon's Bee launched the $74.99 Bee Frontier wearable and announced the $99.99 Bee Eclipse phone recorder and an $11.99 per month Bee Pro plan.
• Kunlun Tech's Skywork AI open-sourced Matrix-Game 4.0, a game engine paired with a video model that it says runs 480p at 60 fps on one RTX 5090.
• Perplexity open-sourced SPACE, its AI agent sandbox, and CoreWeave, Akamai and Crusoe agreed to offer it, Semafor reports.
• Ant Group reportedly released Ling-3.1-flash weights under the MIT license, a 560B model with 25B active parameters, according to Pandaily.
• Shanghai AI Laboratory says mathematicians used Intern-S2 to prove the 3D Ghys conjecture, with the proof verified in Lean, Pandaily reports.