This Week in AI New Models Flood In and Agents Get Safer

New Models Go Wide Open

This week was dominated by companies releasing powerful open-weight models that anyone can download and run. Mistral unveiled Mistral Large 4, a multimodal model with 1.05 trillion total parameters but only 52 billion active ones at a time thanks to its Mixture of Experts design, released as a public preview on October 6. StepFun made its 600-billion-parameter Step 5 Preview available on Hugging Face and ModelScope, while Meta launched Muse Spark Open — an 816-billion-parameter model built to run locally on Apple Mac Studios, though it requires at least one terabyte of RAM. Recursive Superintelligence entered the fray with Flint, a 40-billion-parameter coding agent designed for consumer hardware, and Moonshot AI previewed Kimi K3.1 Pro, a 2.8-trillion-parameter model scoring just three points behind Claude Opus 5.5 on Artificial Analysis at half the cost, with full open weights arriving November 4.

The speed story ran through several releases. Inception Labs introduced Mercury 3, a diffusion-based language model generating complete answers in parallel rather than word by word, reaching roughly 1,068 tokens per second on NVIDIA GPUs. Liquid AI launched RIPL, an engine running smaller models in under 50 milliseconds. Meanwhile, Qwen researchers developed DScout, a technique that can make text generation up to four times faster, and Stanford and UC Berkeley researchers created TokenGate, an add-on that lets models like Qwen3 skip irrelevant text and read up to 11.3 times faster without losing accuracy.

Agents Learn to Be Better Behaved

AI agents got some serious upgrades this week. Anthropic released sandbox-runtime, a free tool that wraps any program in filesystem and network restrictions without needing containers. A Y Combinator startup called Alter built a system giving every AI agent its own identity and limited permissions instead of sharing one powerful credential. Laminar introduced a cheap open-source model that detects hidden errors in agent workflows at 23 times less cost than top commercial alternatives. Researchers also published work on LEASH, a framework letting small agents call bigger models only when unsure, potentially cutting inference costs by 19 percent. A study on agentic reinforcement learning found that diversifying training environments beats grinding out more episodes, boosting compute efficiency by up to 1.7×. And Microsoft Research released Agensh, an open-source system letting up to 1,024 AI coding agents work together without a central coordinator.

Safety, Research and Tools

On the safety side, OpenAI reported that its internal team Calif using GPT-5.6 Sol discovered two critical vulnerabilities in PHP ahead of attackers as part of its Patch the Planet initiative, with Wiz independently confirming both. Anthropic also launched a free OSS Scanner that found 77 critical and 803 high-severity issues across open-source projects in its first 11 days. A new study warned that training safe models on their own outputs through recursive self-distillation can quietly erode safety over time. US DHS banned DeepSeek from all federal civilian agency systems, and DeepSeek itself open-sourced its high-performance inference engine DeepInfer under an Apache 2.0 license.

Research made waves too. MIT researchers built IRIS, an AI system decoding how cells communicate chemically during development, potentially speeding up stem cell engineering. Google's Gemini 4 Argon placed third in the StarSkimmish Bench writing StarCraft bots, and Yongchao Huang introduced MDN-VJEPA, a system predicting multiple branching futures at once instead of one.

In tools, NVIDIA released Structured Data Models using in-context learning for spreadsheet predictions, Baidu launched open-source Unlimited OCR 2 for reading hundreds of pages at once, vLLM hyper-optimized Kimi K3 to over 840 tokens per second, and Anthropic changed how it reviews code after Claude Code started writing 90 percent of its own code, doubling merge rates while cutting reverts by 40 percent. Meanwhile, Anthropic announced that clicking thumbs up or down on Claude can now store full conversations for up to five years for model training.

* *This story reports on a trend for information only.

The week in short

Mistral Large 4: 1.05 trillion parameters, 52 billion active, released October 6
Mercury 3: ~1,068 tokens/second, generates answers in parallel not word by word
DeepSeek banned from US federal civilian agency systems by DHS order
Anthropic OSS Scanner found 77 critical and 803 high-severity issues in 11 days
Kimi K3.1 Pro: 2.8 trillion parameters, full open weights arriving November 4