AI Week in Review 25.11.08

Kimi K2 Thinking, Veo 3 adds camera movement, AI2's OlmoEarth, MAI-Image-1 on Bing, LongCat Flash Omni, Xpeng's IRON humanoid robot, Windsurf Codemaps, Maya1 TTS, File Search Tool in Gemini, any-llm.

Figure 1. Still from Theoretically Media demonstrating Veo 3’s camera movement feature. Google Veo 3’s camera movement feature adds a prompted camera movement to adjust a video, keeping the original scene while changing camera angle on it.

Top Tools

Moonshot AI has announced Kimi K2 Thinking, an open AI reasoning model designed from the ground up for powerful agentic AI. Based on Kimi K2, Kimi K2 Thinking is a highly efficient MoE (mixture-of-experts) model with 1 trillion total parameters and 32 billion active parameters during inference, with a 256,000 token context window. On math, coding and agentic benchmarks, Kimi K2 Thinking performs on the level GPT-5 and both GPT-5 and Claude 4.5 Thinking: 44.9% on Humanities Last Exam (HLE), 71.3% on SWE-bench verified, and 60.2% on BrowseComp (agentic search).

While it has excellent general capabilities in writing, coding and reasoning, it really shines on agentic tasks and tool-calling:

Kimi K2 Thinking can execute up to 200 – 300 sequential tool calls without human interference, reasoning coherently across hundreds of steps to solve complex problems.

Kimi K2 Thinking was trained using Quantization-Aware Training (QAT), so that quantized versions won’t have compromised output quality. This makes the benchmarks even more impressive because benchmark results are reported based on INT4 precision.

This is perhaps the most important AI model out of China since DeepSeek R1, an open-source SOTA AI model from China that competes with GPT-5 just months after the GPT-5 release. It once again shows that Chinese AI labs are closing the gap with leading US-based AI labs.

Figure 2. Kimi K2 Thinking beats all other AI models on key benchmarks. Note that Kimi K2 Thinking requires tool calling to get the 44.9% result on HLE.

AI Tech and Product Releases

Ai2 launched OlmoEarth, an open, end-to-end platform for “Earth intelligence” to help generate insights about our planet. OlmoEarth consists of:

• The OlmoEarth foundation models, small AI models (from 1.4M to 300M parameters) trained on Earth observation data to perform geospatial tasks such as scene classification and semantic segmentation.

• OlmoEarth Studio, which provides dataset modelling and fine-tuning.

• The OlmoEarth Run workflow engine to run earth-modelling processes.

• OlmoEarth Projects, useful fine-tuned example models for Earth intelligence.

Figure 3. OlmoEarth Studio interface, showing how it can process imagery and other geo-spatial information to model Earth features.

Microsoft rolled out its proprietary text-to-image generator MAI-Image-1 to Bing Image Creator and Copilot. MAI-Image-1 provides fast, high-quality image generation of food, nature and real-world scenes, and moves Microsoft away from reliance on OpenAI’s image.

The Meituan LongCat team released LongCat Flash Omni, an open-source omni-modal (text, audio, video input) model, with an MoE architecture with 560B parameters and 27B active parameters. This model achieves SOTA performance for omni-modal AI models, as shown in the LongCat-Flash-Omni Technical Report that describes training and performance.

Chinese EV company Xpeng unveiled the IRON humanoid robot, claiming it has a ‘most human-like’ design with soft skin, bionic muscles. The humanoid bones, muscles, and skin make this the most humanoid of robots yet; it is setting off uncanny-valley and “Sci-Fi is here” alarm bells. They also announced VLA 2.0 large model, a vision-language-action model for physical AI, useful for both self-driving cars and AI robots.

Figure 4. XPeng’s IRON robot is the most humanoid robot yet.

Cognition announced Windsurf Codemaps, which produces navigable maps of your codebase powered by SWE-1.5 and Sonnet 4.5. Codemaps is intended to provide contextualized codebase understanding, grounded in code navigation. Think of it as Notebook LM for codebases.

We all need more AI that turns your brain ON, not OFF. – The Cognition Team

The Maya Research team has released the Maya1 voice generation model, an open source 3B parameter text-to-speech (TTS) model that has climbed to #2 among open-weight voice AIs and #20 globally on quality benchmarks. We are now at the point where high-quality TTS models are freely available. A demo is available here and Maya1 model weights are on HuggingFace.

Google has introduced the File Search Tool in Gemini API, which provides an end-to-end retrieval-augmented generation (RAG) features within Gemini. This handles complex RAG details for developers and streamlines the RAG process. It supports built-in citations, many file formats, and powerful semantic vector search.

Google’s no-code app builder Opal is now available to more than 160 countries worldwide.

Mozilla.ai released any-llm1.0, an open-source interface that lets developers plug in any model, cloud or local into their AI workflow. The any-llm decouples project logic from model provider with a consistent API, letting developers seamlessly plug in any LLM without rewriting integration stacks.

Higgsfield announced Recast, the ability to recast a character in any video with another character.

Sora is now available on Android, making Sora 2’s advanced video generation capabilities available to millions more users. Android users will still need an invite code to actively use Sora’s generation features, and the launch is limited to select countries.

Google’s next image generation model release, aka Nano Banana 2, could be available as soon as next week. Nano Banana became popular thanks to its SOTA features in image editing, and an improved release could significantly boost AI-assisted creative workflows.

AI Research News

Google will launch two satellites carrying its TPU AI chips in 2027, as part of Project Suncatcher, an exploration of space-based AI infrastructure. The project envisions space-based data centers consisting of many satellites, loaded with TPUs and connected by optical inter-satellite links to power distributed AI tasks. Google hopes that in coming decades, space-based data centers will become more cost efficient than land-based, since solar panels are up to 8-times more effective in orbit.

AI Business and Policy

Apple is makng a deal to pay Google $1 billion per year to use a 1.2 trillion Google-based AI model in the next-gen Siri voice assistant. This is yet another shift in Apple’s AI sourcing for Siri. Apple’s failure to update Siri has left them as a notable laggard in the AI race.

OpenAI has published a “Teen Safety Blueprint” aimed at guiding AI products and policies to protect and empower teens in their use of AI. The Blueprint emphasizes age-appropriate design, meaningful product safeguards, and ongoing research and evaluation. OpenAI says it is proactively implementing protections across its products, including parental controls and an age-prediction system, rather than waiting for regulation to dictate AI safety for youth.

OpenAI and Amazon announced a US$38 billion deal, allowing OpenAI to run its models on AWS infrastructure, giving OpenAI access to hundreds of thousands of Nvidia chips on AWS. This continues OpenAI’s push for more access to computing and cloud infrastructure to power ChatGPT inference and training.

Perplexity says Amazon issued a legal threat to block Comet AI assistants from shopping on Amazon. Perplexity is calling it “Bullying,” and notes that Amazon is motivated by a desire to monetize AI agent usage through upselling.

In contrast, Shopify is leaning in on agentic AI’s expanding role in shopping. Shopify says AI traffic is up 7-fold since January, and AI-driven orders are up 11-fold. Shopify is working with Perplexity and Microsoft to add in-chat shopping experiences and partnered with OpenAI on ChatGPT integration.

Anthropic is expanding its global operations by opening new offices in Paris and Munich. Europe is Anthropic’s fastest-growing region, where it has enjoyed over 9-fold revenue growth from businesses in Europe adopting Claude AI. Anthropic is targeting $20-26 billion in annual recurring revenue for 2026.

The Associated Press announced AP Intelligence, a structured-data offering that transforms its verified news journalism into datasets and intelligence feeds for AI and analytics use across industries.

The European Commission is exploring a pause of parts of the EU’s landmark AI Act amid pressure from Big Tech firms and engagement on trade concerns with the U.S. While the law officially entered into force in August 2024, implementation timing for some provisions may be delayed.

AI Opinions and Articles

In a Vanity Fair interview, Kim Kardashian discussed her “frenemy” ChatGPT, blaming it for failing her Law Exam. She revealed using ChatGPT for legal advice but claims its “hallucinations” of false information caused her to fail law tests. If LLMs generate fake answers, leading to real-world problems for users and even legal professionals.

The Collins Dictionary Word of the Year 2025 is “vibe coding,” describing the use of natural language AI prompts to write computer code and generate AI applications. AI changes everything, including even language.