The Day in AI

• MiniMax released M3.1 Flash, an agentic and coding model priced at $0.15 per million input tokens, which Artificial Analysis scores 42 on its Intelligence Index.
• Amazon Web Services released TorchNeuron in public beta, a PyTorch backend that runs existing PyTorch code on Trainium chips after a one-line configuration change.
• Inception Labs released Mercury 3, a diffusion language model it says runs at 1,068 tokens per second with a 1 million token context window.
• VAST open-sourced EdenLive, a real-time renderer that turns coarse game footage into styled frames above 24 frames per second on one RTX 5090.
• Disney is set to hire more than 60 Character.AI employees, including former CEO Karandeep Anand, according to Ars Technica and Business Insider.

MiniMax says M3.1 Flash is nearly three times faster than M3 and costs 70% less per agentic task. MiniMax credits more aggressive sparsity for the speed, so fewer parameters are active per token, and reports output above 200 tokens per second with a median of more than 170. Before launch the model was tested anonymously as Space Bunny, which MiniMax says processed more than 200 trillion tokens. Cached input costs $0.015 per million tokens, and the model is available only through the API.

The benchmark claims are company-reported. MiniMax gives 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 82% on CyberGym. The independent numbers come from Artificial Analysis, as reported by The Decoder, which puts the cost at about $0.22 per task and the hallucination rate at 93%. That hallucination figure is high, and it's the number to check before using the model for anything factual.

Amazon Web Services released TorchNeuron in public beta on 19 October. It's a PyTorch backend adapted from PyTorch itself, so developers can move existing PyTorch apps onto Trainium without rewriting them. Amazon says 85% of AI apps are written in PyTorch and that it is sending changes upstream to remove code that ties the framework to specific chips. Amazon cites Deepgram as processing 70% more audio per dollar after moving to Trainium.

Inception Labs released Mercury 3 on the same day. A diffusion language model starts from a rough draft of the whole output and refines many tokens in parallel over a few steps, and this lets it generate text faster than models that write one token at a time. Inception says Mercury 3 matches GPT-6 Luna and Claude Haiku 5.5 on benchmarks while running 4 to 8 times faster on NVIDIA GPUs. It has more than 40 times the parameters of Mercury 1, and Inception presents this as evidence that scaling laws hold for text diffusion. The list price is $0.20 per million input tokens and $0.70 per million output tokens, halved at launch. None of the speed or quality figures have been independently measured yet.

• VAST open-sourced EdenLive in 5B and 14B sizes, which renders footage from a game engine that keeps running the game logic, guided by a first frame and a text prompt.
• Disney is in a deal to hire most of Character.AI's staff after the company filed for bankruptcy, according to Ars Technica citing court filings, and a judge must approve it.
• US Department of Homeland Security plans to ban DeepSeek from federal civilian agencies, according to CyberScoop.
• DeepSeek is working with Huawei on future chip designs and plans to pretrain on Ascend 950DT as soon as the end of 2026, according to The Information.
• Nscale is in talks on a second deal with Anthropic worth about $45 billion for 460 megawatts in West Virginia on Vera Rubin chips, according to Bloomberg.
• Meta described Prometheus, its gigawatt-scale Ohio supercluster that trained the Muse models and will have about 1 million GPUs at full scale.
• NVIDIA released Nemotron 3.5 Super VL on 15 October, an open 120B multimodal mixture-of-experts model with 12B active parameters and a 1 million token context.