NVIDIA NVFP4 Breakthrough, Local Model Milestone, and Zepto IPO Lead This Week's AI and Business News

Key Takeaways

1. NVIDIA achieves 1.31-1.73x training speedup with NVFP4 precision on Blackwell, zero accuracy loss, marking a leap in low-precision training.
2. Local models reach 71.3% accuracy on real-world queries; consensus shifts toward multi-model inference over reliance on frontier models.
3. Hugging Face releases Mellum2 (MoE) and Qwen3.5 quantized checkpoints; Arcee AI fully migrates from AWS S3 to HF, underscoring open-source ecosystem growth.
4. Apple expands Private Cloud Compute to Google Cloud using NVIDIA GPUs, deepening AI infrastructure collaboration.
5. Zepto files for India IPO at $7-10B valuation, with $2.4B revenue and halved losses, validating 10-minute delivery model.
6. Meta AI sees 2.5x user growth but only 4.5% 30-day retention, raising questions about organic acquisition.

1. AI Models and Infrastructure

1. NVIDIA demonstrated training Llama 3 8B and 405B on Blackwell using NVFP4, achieving 1.31-1.73x speedup over FP8 with zero accuracy loss. — via 1
2. Yann LeCun introduced VLA-JEPA, which combines JEPA world models with action dynamics; it runs in real-time after fine-tuning on just 13 examples and is open-sourced on LeRobot. — via 1 2
3. Hugging Face released Mellum2 (MoE architecture, low latency, high throughput) and Qwen3.5 quantized checkpoints optimized for Apple hardware; Qwen3.7 is expected this week. — via 1 2 3
4. Both Hugging Face and Yann LeCun cite Stanford research showing local models achieve 71.3% accuracy on real-world queries, reinforcing the trend that most tasks do not require frontier models. — via 1 2
5. Hugging Face predicts that by 2026, AI agents will primarily run on the command line, and has updated its hf CLI to support agentic language; model routing technology is growing rapidly. — via 1

2. Enterprise and Market Dynamics

1. Zepto filed for an India IPO at an estimated $7-10 billion valuation. The company doubled revenue to ~$2.4 billion last year, halved losses to -26%, and has nearly 50 million users with growing per-user revenue. — via 1
2. Meta AI grew users 2.5x but retains only 4.5% at 30 days, lagging other AI apps. Deedy suggests growth may not be organic. — via 1
3. Perplexity CEO announced Quartr's native MCP connector on Perplexity Computer (43 tools, 15,000+ company IR data), the Billion Dollar Build competition with 1,500 teams and finals on June 9, and the Billion Pound Build competition offering £1M in computing credits. — via 1 2 3
4. Apple is expanding Private Cloud Compute to Google Cloud using NVIDIA GPUs, marking a deeper collaboration in private AI. — via 1
5. Arcee AI became the first AI lab to fully replace AWS S3 with Hugging Face for model and dataset storage, backed by a multi-million dollar deal to support US open-source AI. — via 1
6. The Rundown received a strategic investment from Electrify (backers of Veritasium, Fireship) to accelerate its mission of educating 1 billion people on AI. — via 1

3. AI Benchmarking and Coding

1. swyx introduced FrontierCode, a new benchmark harder than SWEBench that measures code maintainability (Opus 4.8 scores 13.8%). He notes AI coding ability has improved significantly since late 2025, reducing required retries from 6 to 2, and predicts each FrontierCode level will be conquered annually. — via 1
2. Anthropic observes that AI progress in coding outpaces biology because biological databases are not designed for agents; building agent-ready infrastructure is crucial. — via 1

4. Controversies and Outlook

1. Ethan Mollick reports that both Anthropic and OpenAI have mentioned possibly slowing AI development, but stress that global coordination is required and no clear method exists yet. — via 1
2. Yann LeCun opposes an AI pause, calling it an "mass hallucination" based on imagined problems that would cause more harm than good. He also cites a Wharton paper claiming AI must boost productivity by 2.7x to avoid tech bankruptcies, with OpenAI reportedly negotiating a bailout with government. — via 1 2