On April 28, Xiaomi open-sourced the MiMo-V2.5 series of large models under the MIT license. The flagship MiMo-V2.5-Pro adopts a MoE architecture with 1.02 trillion total parameters and 42 billion activated parameters, supporting up to 1 million tokens of context. It outperforms DeepSeek-V4-Pro and Kimi K2.6 on benchmarks such as GDPVal-AA and Claw-Eval. In ClawEval tests, it consumes only about 70,000 tokens, saving 40%-60% compared to competitors. On its first day of release, it completed integration with seven chipmakers including Alibaba's Pingtouge, AMD, and Baidu's Kunlun芯, and supports inference frameworks SGLang and vLLM. Concurrently, Xiaomi launched a free incentive program offering up to one hundred trillion tokens and an Agent ecosystem co-construction initiative.

NVIDIA has released Nemotron 3 Nano Omni, featuring a 30B-A3B hybrid Mamba-Transformer-MoE architecture that unifies vision, audio, and language processing, with support for up to 256k tokens of context. It leads in multimodal benchmarks such as MMLongBench-Doc, OCRBenchV2, and WorldSense, achieving up to 9.2x higher throughput than comparable open-source full-modal models. The model weights, dataset, and training methodology are fully open-sourced, supporting deployment from edge devices to the cloud. Enterprises including H Company, Palantir, and Foxconn have already adopted it. It is available via platforms such as Hugging Face, Ollama, and OpenRouter.

On April 28, OpenAI and AWS announced an expanded strategic partnership, bringing cutting-edge models like GPT-5.5, the Codex programming agent, and Amazon Bedrock Managed Agents to the AWS platform. Enterprises can now leverage OpenAI capabilities within AWS's existing security, compliance, and procurement workflows. Codex supports code generation, testing, and documentation processing, while Managed Agents can execute multi-step tasks with persistent context. Usage counts toward existing AWS cloud spending commitments. This move marks OpenAI breaking its exclusive cloud partnership with Microsoft and officially entering the multi-cloud era.

Photonic computing company Lightelligence debuted on the Hong Kong Stock Exchange on April 28 at an issue price of HK$183.2, opening at HK$880—a surge of over 380%—with a closing market capitalization of HK$81.5 billion. The IPO raised approximately HK$2.5 billion, with the retail portion oversubscribed by 5,784.7 times. Founded by MIT PhD Shen Yichen, the company holds three core technologies including photonic matrix computing, generating RMB 106 million in revenue in 2025. Cornerstone investors include Alibaba, Tencent, Temasek, and GIC. China’s Scale-up optical interconnect market is projected to reach RMB 180.5 billion by 2030.

Ineffable Intelligence, a UK-based AI startup founded by former Google DeepMind lead researcher David Silver, has secured $1.1 billion in seed funding, valuing the company at $5.1 billion—the largest seed round ever for a European startup. Investors include Sequoia, Lightspeed, NVIDIA, Google, and the UK government's sovereign AI fund. The company aims to develop 'super learners'—AI systems capable of autonomous learning through self-experience without human data, leveraging reinforcement learning technology toward achieving superintelligence. The British Business Bank has additionally invested $20 million.

On April 27, China's National Development and Reform Commission (NDRC) ordered Meta to reverse its $2 billion acquisition of AI startup Manus, despite the transaction being completed and technical and personnel integration already achieved. Although Manus had relocated from Beijing to Singapore, its founders and core technology originated in China, placing it under Chinese jurisdiction. This sets a precedent for forced 'reverse mergers,' signaling that strategies to circumvent regulations via offshore registration are no longer effective. Meta now faces the challenge of disentangling integrated code and employees, reshaping regulatory risks for cross-border M&A involving Chinese AI firms.

On April 28, the Chinese Academy of Sciences (CAS) launched the 'Panshi 100' model system, building a cluster of eight discipline-specific large models in mathematics, physics, materials, astronomy, environment, aerospace, geography, and life sciences, anchored on a foundational scientific LLM. The system has aggregated over 2,000 research tools, serving more than 50 institutions and over 100 research scenarios. The 'Digital Cell' model successfully identified and validated three new drug targets within 30 days; 'Jinwu' achieved intelligent solar flare prediction; 'Yuheng' constructed a global carbon emissions map. This marks a shift toward platformized innovation in AI for Science.

On April 28, IBM globally launched IBM Bob, covering the entire software development lifecycle—planning, coding, testing, deployment, and modernization—with integrated governance, compliance, and security controls. Over 80,000 IBM employees have used it, with respondents reporting an average 45% increase in productivity. Bob supports multi-model orchestration (Claude, Mistral, Granite, etc.), intelligently routing tasks across models. Blue Pearl reduced a 30-day Java upgrade to just 3 days; APIS IT achieved 10x faster architecture analysis. Now offered as a SaaS solution with a 30-day free trial.

On April 28, MIIT and the National Data Bureau jointly issued a notice launching the 2026 'Model-Data Resonance' initiative, focusing on key industries including steel, automotive, aerospace, and medical equipment to promote a virtuous cycle among 'data, models, and application scenarios.' On the same day, the Political Bureau of the CPC Central Committee proposed a comprehensive rollout of the 'AI+' initiative to foster a new form of intelligent economy. The Cyberspace Administration of China held talks with Jianying, Maobox, and Jimeng AI, requiring compliance with AI-generated content labeling regulations. A wave of policy actions reflects China's dual focus on AI governance and industrialization.

Anthropic's valuation has surpassed $1 trillion in secondary market trading, making it the third privately held company—after OpenAI and SpaceX—to achieve this milestone. Data comes from secondary platforms such as Jupiter's Prestocks and Forge Global. The company's annualized revenue is expected to exceed $30 billion, driven largely by widespread enterprise adoption of Claude Code. It previously closed a Series G round at a $380 billion valuation in February, with Google committed to invest up to $40 billion. The market widely expects Anthropic to pursue an IPO in 2026.