Japan’s Digital Agency opens tender for 100% domestic LLM to run all gov’t services by 2026 major step for AI sovereignty as global race intensifies. Deadline Jan 2026.

DeepSeek launches V3.2-Speciale: open-source Chinese LLM that claims to beat GPT-5 in reasoning & match Gemini 3.0 Pro. New king challenging US dominance

Tether drops QVAC Fabric—its first LLM built for blockchain & crypto. Targets secure, decentralized AI processing in finance. Stablecoin giant joins the race

OpenAI’s internal “Garlic” model already outperforms Gemini 3 & Claude Opus 4.5 in coding/reasoning. Company declares “state of emergency” against Google surge

MIT unveils SEAL: breakthrough technique lets LLMs permanently learn new info by auto-generating “study sheets,” ending the era of static knowledge models.

Google rolled it out globally via Gemini 3 Pro (Search, Gemini app, Workspace, API) and it now turns text, tables or notes into accurate infographics, diagrams, or 4K-ready visuals — with crisp multilingual text, layout control, and consistent character rendering.

AWS unveils serverless tools in SageMaker & Bedrock to simplify building tailored LLMs with natural language prompts. Includes $100K credits for Nova models. Eases enterprise AI dev, intensifying customization race

Mistral AI launches 675B-param Large 3 (41B active MoE) outperforming DeepSeek on benchmarks, plus 9 Ministral variants (3B-14B) for edge use. All Apache 2.0 open-source, multimodal w/ 256k context.

Digital Agency opens bids for Japanese-made LLMs/SLMs tailored for admin tasks like chat & public services. Testing on "Genbu" platform in FY2026; aims to embed in gov AI apps.

OpenAI preps new LLM codenamed Garlic amid ChatGPT's "code red" push for advanced coding. Signals next-gen dev tools, but details sparse—expect boosts in reasoning & efficiency. (132 chars)b692ba968c8d

OpenAI trains models to "confess" cheating/lying for better internal transparency & diagnostics.

MIT's instance-adaptive scaling lets LLMs cut compute ~50% on tough reasoning tasks.

Zhipu launches GLM-4.6V: multimodal with native tool-calling & 128K context.

Chinese open models (DeepSeek, Qwen) dominate adoption surge; Mistral/Nvidia partnership deepens.

AWS re:Invent boosts custom LLM tools in Bedrock/SageMaker for easier enterprise fine-tuning.