The Day in AI

• DeepSeek released Super-agent on 31 October, an open-source research prototype that watches developer chats and proposes code fixes without being prompted.
• SenseTime open-sourced SenseNova-Loop-9B-Preview, a 9B looped Transformer that runs part of its layer stack twice and that SenseTime says matches models of about 27B.
• Microsoft published Mage-VL-Cowork, a 4B computer-use agent under Apache-2.0, and reports 66.0% on OSWorld-Verified with a 50-step budget.
• ModelBest and OpenBMB open-sourced the MiniCPM-OCR series, a 1.3B document parser and a 2B model that scores OCR output quality.
• Amazon reportedly said on its Q3 earnings call that Bedrock now offers Moonshot's Kimi K3 and Z.ai's GLM 5.3 open models.

Super-agent uses a router and worker design. A router collects input from chat and email and passes work to worker agents, which act like a small virtual team. DeepSeek's reasoning is that one person can supervise only three to five agent sessions at once, so the system handles conversation management and context sync while humans set goals and review results. The public version works with a single GitHub repository and can also take input through Feishu. Arageek reports tasks finishing in under 16 minutes against over 29 minutes for a traditional approach, but that figure doesn't appear on a primary DeepSeek page.

SenseTime open-sourced SenseNova-Loop-9B-Preview on 30 October. A looped Transformer runs some of its layers more than once with the same weights, which adds compute per token without adding stored parameters. This model stores 32 decoder layers and runs layers 4 to 27 twice, for 56 layer executions per forward pass and about 14.6B parameters at inference from about 9B stored. SenseTime also adds two-stream manifold-constrained hyper-connections (mHC) inside the looped section, and in its own ablations mHC added 4.42 points on GPQA Diamond and 13.34 points on TeamBench.

Microsoft published Mage-VL-Cowork on 30 October, a 4B vision-language model that controls a GUI and runs terminal commands from one set of weights. Microsoft reports 66.0% on OSWorld-Verified with a 50-step budget and 87.2% on ScreenSpot-v2. Training ran in five stages and ended with online reinforcement learning (RL) using Group Relative Policy Optimization (GRPO). Microsoft says the RL stage cut output tokens by 25.4% and model calls by 13% compared with supervised fine-tuning alone.

• ModelBest and OpenBMB released MiniCPM-OCR 1.0 (1.3B) for document parsing and MiniCPM-OCR Jev (2B), which compares a source image with OCR output and assigns confidence scores. It follows Xiaomi-OCR-0 the day before.
• Amazon CEO Andy Jassy listed Kimi and GLM models on Bedrock during the Q3 call on 29 October. The exact versions Kimi K3 and GLM 5.3 aren't confirmed in the visible source.
• Google said on 31 October that AppFunctions, its Android interface for exposing app functions to agents such as Gemini, is in beta, with the Gemini integration reportedly still in private preview.
• Alibaba will reportedly integrate its Qwen model into Apple Intelligence, according to CGTN.
• Lumaril, which builds software for industrial production control, raised a $66M Series A at a $400M valuation, according to the Wall Street Journal.
• Google reportedly joined OpenAI and Anthropic in supporting a robots.txt signal for AI training preferences, according to TechloMedia.
• OpenAI is merging ChatGPT and Codex usage limits, according to its help centre.
• Westlake University opened a world model research centre led by Yu Kaicheng, Pandaily reports.