NVIDIA Blackwell Takes Top Spot in AgentPerf, the First Agentic AI Benchmark
In the inaugural results of AgentPerf, the industry's first agentic AI benchmark announced by Artificial Analysis, NVIDIA's Blackwell Ultra NVL72 (GB300 NVL72) platform delivered overwhelming performance. Compared to NVIDIA Hopper (HGX H200), it can run up to 20x more agents concurrently per megawatt.
Agentic AI is a fundamentally different workload from traditional chat completion. While chat finishes with a single LLM call, agents break goals down into multiple steps and chain together dozens to hundreds of tool calls — compiling and running code, querying databases, browsing the web, and more. Existing AI inference benchmarks only measured a single LLM call and failed to capture this agent-specific load.
AgentPerf is built on real coding agent trajectories, reproducing realistic workflows across more than 12 programming languages, including file reads, code edits, command execution, and iterating based on results.
In a test using DeepSeek V4 Pro, a frontier-class MoE model, the GB300 NVL72 achieved these results by connecting 72 GPUs into a single-rack system, with full-stack optimizations that streamline the execution of large-scale MoE models. Inference providers such as Baseten, DeepInfra, and Together AI are already offering agentic workloads on Blackwell, which are reportedly being used in production products like Cursor and Pam.ai.
NVIDIA's Vera Rubin architecture has also already entered mass production, and the company says next-generation infrastructure to meet growing agentic AI demand is on track.
Meta's AI Division Faces Internal Collapse Amid a 'Soul-Crushing' Environment
Reports from TechCrunch and WIRED reveal that Meta's AI division, only months after its founding, is grappling with serious internal problems. According to whistleblowers and people familiar with the matter, the 6,500-person organization suffers from working conditions so grim that they have been described as a 'soul-crushing gulag.'
WIRED reports that both executives and employees are struggling with Meta's chaotic AI strategy. Internal backlash against Mark Zuckerberg's direction has surfaced, and at one internal meeting, employees reportedly confronted leadership in unusually blunt terms.
TechCrunch notes that the organization is 'on the verge of revolt.' The picture that emerges is one of frontline engineers deeply frustrated with the management structure and strategic direction of the rapidly expanded AI division.
While Meta is pouring massive investment into AI and racing to develop LLMs and roll out AI-integrated features, these reports suggest the company faces serious challenges in how the organization itself is run.
Sparse Attention Support Still Missing in MiniMax M3's GGUF Build
The Reddit LocalLLaMA community is debating a technical issue with the GGUF release of MiniMax M3. The Hugging Face page for unsloth/MiniMax-M3-GGUF carries a note stating that 'MiniMax's Sparse Attention is not yet supported, and inference falls back to Dense Attention,' which means the 428B-parameter model uses all of its weights at every step — potentially making it significantly slower than when its signature sparse attention is leveraged.
MiniMax M3, recently released with open weights, is a 427B-parameter natively multimodal model, and its sparse attention-driven inference efficiency was one of its headline features. When run locally in GGUF format, however, this optimization does not apply, so you effectively pay the full computational cost of the dense model.
The community is discussing 'just how much slower it is compared to the full implementation,' and until sparse attention support lands, practical local use remains a tall order.
Claude Code v2.1.173–v2.1.176 Ship in Quick Succession — Model Governance and Remote Control Fixes Take Center Stage
A Qiita post translates and breaks down the official Claude Code changelog, summarizing changes across the four releases from v2.1.173 to v2.1.176.
Rather than new features, these back-to-back releases center on 'model governance' and fixes around background execution and Remote Control — foundational hardening that is said to be key to running Claude Code reliably as a CLI tool.
Claude Code is Anthropic's AI coding assistant, rapidly gaining adoption as an agentic development tool that runs in the terminal. Improved stability for background execution and remote control is a particularly meaningful improvement for embedding it into CI/CD pipelines and automating long-running tasks.
Visa × OpenAI Partnership — Ushering in the Era of AI Agents with Wallets
At the Visa Payments Forum held in San Francisco on June 10, 2026, Visa and OpenAI announced a strategic partnership. According to a Qiita explainer, this marks the turning point from the 'ask AI to shop' stage to the 'AI completes the payment' stage.
The two companies will jointly build infrastructure that lets AI agents actually execute card payments on users' behalf, filling in a critical piece of the puzzle for making agentic AI practical.
While the partnership itself was already covered in the news on June 11, the Qiita article also details system design and implementation considerations from an engineer's perspective, making it well worth a read from a technical standpoint as well.