Key AI Launches This Week
AWS Launches Nova Forge and Custom Model IP
Amazon Web Services (AWS) unveiled Nova Forge, an "open training" platform that offers enterprises deep control over AI model assets. Along with the launch, AWS expanded its Nova 2 model family. The platform shifts from the previous practice of offering AI models through an API as a third-party service, allowing customers to co-create IP by blending proprietary data into foundational AI models at various training stages. Nova Forge also introduces "Novellas," proprietary models that combine Nova's full reasoning power with specific business knowledge. The platform features Reinforcement Learning "Gyms" for agent training and Synthetic Data Distillation to create cost-efficient models. Enterprises can now maximize the competitive advantage from their unique data, achieving faster ROI with customized models like Nova 2 Lite.
Read more: AWS Nova Forge Announcement Additional AWS Updates
DeepSeek-V3.2 Update
DeepSeek has released version V3.2, alongside a high-compute variant, DeepSeek-V3.2-Speciale, designed to surpass GPT-5 in reasoning tasks and match Google Gemini 3.0 Pro. The new version integrates tool-use thinking, enabling the model to pause its chain of thought, execute external tasks like running scripts, and continue reasoning. The DeepSeek Sparse Attention (DSA) architecture improves efficiency in handling long contexts. DeepSeek-V3.2-Speciale also provides a cost-effective solution, outperforming GPT-5 while offering a price reduction by 10x. It is ideal for complex agentic tasks, such as analyzing large datasets or running multiple self-verification loops.
Read more: DeepSeek-V3.2 Technical Report API Update
Google's Titans + MIRAS: AI Long-Term Memory Breakthrough
Google Research has introduced the Titans architecture and the MIRAS framework, enabling AI models to maintain genuine long-term memory. This system replaces static memory vectors with a deep neural memory module, enabling continuous learning during inference. The Titans approach allows AI to adapt and prioritize information in real time, offering improvements in extreme-scale context handling, speed, and accuracy. With over 2 million tokens of context, the new models can excel at tasks requiring continuous learning, such as genomic analysis or time-series forecasting.
Read more: Titans + MIRAS Announcement
AWS re:Invent 2025 + Open-Source AI Updates
At AWS re:Invent 2025, the company announced the launch of new models, agent frameworks, and next-gen hardware. Nova 2 Omni, a multimodal foundation model, and Frontier Agents—autonomous AI agents—were introduced. These updates mark AWS's shift from hosting models to offering comprehensive AI agent development platforms. Additionally, DeepSeek V3.2’s Sparse Attention architecture is highlighted for reducing compute costs while enabling efficient long-context inference. AWS’s hardware advancements with EC2 Trn3 UltraServers bring significant improvements in speed and energy efficiency, further supporting their move to enhance enterprise AI capabilities.
Read more: AWS re:Invent 2025 Highlights
Anthropic Acquires "Bun" Team: Strategic Move for Claude Code
Anthropic has acquired the team behind Bun, the ultra-fast JavaScript runtime, following the success of their Claude Code tool. This acquisition marks Anthropic's shift toward vertical integration, owning both the intelligence layer (Claude) and execution layer (Bun). By leveraging Bun’s capabilities, Claude can generate self-contained tools and improve the "write-run-fix" cycle, enabling faster and more reliable autonomous software generation. The acquisition also helps ensure that future optimizations in runtime speed and reliability will be tightly integrated with Claude.
Read more: Anthropic Acquires Bun Announcement
Mistral AI Releases Mistral 3 Models for Open-Source Edge and Cloud
Mistral AI has launched its Mistral 3 family, which includes a mixture-of-experts model and three efficient edge models. The models are released under the open Apache 2.0 license, emphasizing the trend toward open-source AI for both edge and cloud applications.
Read more: Mistral 3 AI Models Release
