Xiaoliu BOT
中文EN

X Platform June 1 AI Brief | NVIDIA Launches Three Major Products, MiniMax Unveils M3 Model, VAST Secures Nearly $200 Million in Funding

NVIDIA Launches Nemotron 3 Ultra, RTX Spark Superchip, and Cosmos 3 at Computex

NVIDIA unveiled three major products at Computex 2026 in Taipei. The Nemotron 3 Ultra features 550B parameters (55B activated), making it the most intelligent open-weight model in the US, scoring 48 on the Artificial Analysis Intelligence Index, surpassing Gemma 4 31B (39) and gpt-oss-120b (33), but trailing Kimi K2.6 (54); inference speed exceeds 300 token/s, far above the 50-100 token/s of comparable Chinese models. The RTX Spark superchip delivers 1 PFLOPS AI compute and up to 128GB unified memory, built in collaboration with Microsoft to create a native Windows Agent runtime environment, planned for launch in autumn. Cosmos 3 is a physical world AI model that unifies physics reasoning, video generation, and action generation into a single architecture, open-sourcing 8B and 32B models along with 6 datasets. Additionally, NVIDIA partnered with OpenClaw to open-source a security scanning dataset of 67,453 ClawHub skills, and Nous Research’s Hermes Agent announced native Windows support integrated into RTX Spark.

Sources:

MiniMax Releases M3 Model: Million-Token Context, Sparse Attention, Native Multimodality

MiniMax released its new flagship model, MiniMax M3, aligning three core capabilities: standard 1M ultra-long context, a new MSA (MoE with Segment-wise Attention) sparse attention architecture, and native multimodal capabilities (text, image, video, desktop operations) integrated from training. MSA reduces per-token computation to about 1/20 of the previous generation under 1M context, with prefill speedup over 9x and decode speedup over 15x. In coding and agent capabilities, it achieves 59.0% on SWE-Bench Pro and 66.0% on Terminal Bench 2.1. API pricing: calls under 512k tokens are 50% off for a limited 7-day period. LufzzLiz commented that this model is a clear bet on agentic coding, highlighting its bundling of long context, tool calling, multimodal understanding, and sustained execution.

Sources:

VAST Completes Nearly $200 Million Funding, Valuation Reaches $1 Billion

3D modeling startup VAST completed A+ and A++ funding rounds totaling nearly $200 million, achieving a valuation of $1 billion and becoming the latest Chinese AI company to join the unicorn club. Lead investors include Ince Capital and China Life Yangtze River Delta Science and Technology Innovation Fund, with follow-on investors including Shenzhen AI Terminal Industry Fund (with Honor as the industry partner), Shenzhen Capital Group, and Yuan Sheng Capital. Existing shareholders such as Primavera Venture Capital and BV Baidu Ventures also made oversubscribed follow-on investments. This marks VAST’s second capital injection in two months after its March funding round. According to Bloomberg, VAST was founded by a 29-year-old gamer.

Sources:

Claude Opus 4.8 Receives Polarized Reviews; Multiple Bloggers Share Hands-On Experiences

Anthropic’s Claude Opus 4.8 has sparked controversy in the community. MaxForAI cited @istdrc’s view that “Opus 4.8 has too many hallucinations,” bluntly stating that only those working on agents know how bad 4.8 is. However, Baoyu (@dotey) offered a different perspective: Opus 4.8 significantly outperforms GPT-5.5 in UI design and implementation, and also excels in system design and planning quality. He recommends using Claude Design for initial design, then implementing with both models to compare differences. He also suggests importing mature Design Systems like Adobe Spectrum 2 to improve consistency. Cell (@cellinlab) used Opus 4.8 to generate a highly polished web-based beach castle game from a game screenshot in under 2 minutes, with impressive physics simulation. Multiple bloggers noted that Claude and Codex subscription quotas have been gradually tightened from 150%.

Sources:

Coze 3.0 Released, Supports Local Agent Integration and Multi-Agent Collaboration

Ma Dongxi NLP (@dongxi_nlp) shared detailed practices of multi-agent collaboration with Coze 3.0. The highlight is the local agent integration feature: a single command can bring local Codex and Claude Code into Coze’s multi-agent team without additional gateway configuration. In a real task, he used three agents to produce a 21-page tutorial—the Codex Agent deeply understood the code repository, the Claude Agent read articles to extract core concepts, and the Coze Agent integrated everything in the cloud to produce an Apple HIG-style HTML tutorial. The key lesson: “set rules first, then assign tasks”—define responsibilities, fix workspaces, agree on notification methods, use project files as the single source of truth, and maintain version traceability.

Sources:

Vibe Coding Tool Ecosystem Continues to Explode: Skill Open-Sourced, Codepilot Refactored, Style Atlas Launched

Multiple bloggers open-sourced a batch of Vibe Coding tools on Children’s Day. Xiangyang Qiaomu (@vista8) announced the free open-sourcing of all recent vibe coding tools and Skills, sharing an AI reading methodology based on Feishu CLI—using Codex to write Epub chapters into Feishu documents, manually annotating and commenting, then having AI interpret, along with a vocabulary learning system (based on the Ebbinghaus forgetting curve) developed with the immersive translation plugin read-frog. Guizang (@op7418) released CodePilot 0.55.0 refactored version, with a complete UI overhaul and support for using Codex as an Agent engine. Gorden Sun open-sourced a Skill for one-click generation of visual math explanation videos. Guyi (@MANISH1027512) retweeted the launch of the Style Atlas on VSC community, covering 620 styles across 10 major style families including materials, photography techniques, animation, and manga, positioned as an aesthetic navigation system for AIGC visual creators.

Sources:

OpenAI Voice Hack Night Showcases Agentic Mobile OS Prototype

Xiaohu (@xiaohu) shared a live demo from a team at OpenAI’s Voice Hack Night: an “agentic operating system” for phones. The core idea is “UI as system”—the phone has no traditional apps; the interface is generated on the fly by an on-device local model, with heavy reasoning offloaded to cloud GPT. The developer used voice commands throughout to book flights, delete calendar events, search AI news, send emails, and create to-do lists. Xiaohu believes this is the AI assistant form everyone dreams of, and a new form that will disrupt the mobile business model, because all interfaces are generated instantly without invoking any app interfaces, challenging Apple’s App Store business model. The demo also encountered a glitch (email sending failed due to login configuration).

Sources:

AI Animation and 3D Cultural Tourism Works Enter Mainstream Platforms

Derek Wen (@derek_wall90176) shared an animated series explaining Chinese idiom stories, which he created using AI in his spare time last year. It is now available on Mango TV for direct streaming. He noted the huge difference in tool effectiveness between last year and this year, with significant improvements in efficiency and quality. He also previewed that his studio’s 3D-style IP series “Huobao Archives,” themed on traditional Chinese culture and history, is planned for release next month. Additionally, @berryxia used Claude to develop a roamable 3D world of Tang Dynasty Chang’an over 2 weeks and $800, supporting real-time voice AI interaction, allowing players to converse with NPCs, enter exhibition halls, and play poetry mini-games. LufzzLiz detailed the technical architecture: the main project uses Three.js for 3D scenes, and the voice sub-project uses Next.js + FastAPI integrated with Agora ConvoAI.

Sources:

Unitree Robotics IPO Approved on STAR Market; OpenAI Recruits Robotics Engineers

According to 36Kr, Unitree Robotics has passed its IPO review on the STAR Market (Shanghai Stock Exchange’s Sci-Tech Innovation Board). The leading quadruped robot company is about to list on A-shares. Meanwhile, Sam Altman posted recruitment information for OpenAI Robotics, stating that the team is developing robots useful to society, with a short-term goal of supporting skilled workers in building infrastructure and a long-term vision of providing personal robots for everyone. OpenAI’s world simulation research project has evolved into OpenAI Robotics, emphasizing co-design of hardware and ML research.

Sources:

Benedict Evans and Lenny Rachitsky Deep Dive: AI Is at the 1997 Stage

Lenny Rachitsky released a podcast conversation with independent analyst Benedict Evans. Evans offered several core judgments: AI is currently equivalent to the internet in 1997—equally significant, but equally early, with most things not yet working well, and the most important use cases likely invisible today; foundation model companies won’t have lasting pricing power, with value aggregating upward; distribution capability is becoming a more important moat than the product itself, as AI makes software easier to build, leading to a noisier market; the logic behind OpenAI and Anthropic acquiring consulting firms is that enterprises lack internal teams to redesign workflows. On employment, he used the history of accounting to illustrate that automation often increases rather than decreases jobs (Jevons paradox), arguing that the key question is not what percentage of your job AI can do, but whether your occupation is a “task” or a “job.”

Sources:

Statistics: Timeline scan lines=360 Bloggers hit=36 Total tweets hit=181 Weighted tweet score=150.25 Original tweets=94 RT tweets=27 Fetch attempts=2 Boundary coverage status=tail_confidently_crossed_target_boundary