NVIDIA releases Nemotron 3.5 Lightning, Gemini crosses 1B users, Lovable raises $400M

Key Takeaways

1. NVIDIA's Nemotron 3.5 Lightning is open-sourced and live on Perplexity's Agent API at $0.0115/$0.17 per million tokens, delivering a high-throughput execution layer for always-on agents. — via 1 2
2. Gemini surpassed 1 billion monthly active users, becoming Google's 14th product to reach that scale and its fastest-growing app. — via 1
3. Lovable raised $400M at a $13.3B valuation, with revenue run rate above $500M (up 7x YoY) and 900M monthly visits, as Deedy argues its real-revenue users counter the "no moat" thesis. — via 1 2
4. Open models advance: LTX-2.5 world model launches on Runway, while MOSS-VL brings image/video understanding to 24GB consumer GPUs via FP8/NF4 quantization. — via 1 2 3
5. New research highlights hidden-reasoning extraction via API exploits and math breakthroughs, while experts debate how to measure AI leadership. — via 1 2

1. Model Releases and Infrastructure

1. NVIDIA's Nemotron 3.5 Lightning, an open-weight MoE model with 30B parameters and only 3B active, is optimized for local execution on laptops or DGX Spark. The release includes weights, data, and training recipes for broad customization. It is positioned as the execution layer for tool calling, verification, and sub-agent tasks. — via 1 2 3
2. Nemotron 3.5 Lightning is now live on Perplexity's Agent API at $0.0115 per million input tokens and $0.17 per million output tokens. It can be paired with frontier models like Nemotron Ultra for planning, with Lightning handling high-concurrency workloads. — via 1 2
3. LTX-2.5 world model is released as an open foundation model for film, robotics, and real-time workflows, featuring improved pixel fidelity and multi-shot consistency. It is now available on Runway, and introduces Diffusion Fidelity Rendering. — via 1 2
4. MOSS-VL's new version adds FP8 and NF4 quantization, enabling local image, video, and real-time streaming understanding on GPUs with 24GB VRAM, with performance close to BF16 and significantly lower memory usage. — via 1
5. Grok 4.6, trained and run on NVIDIA GB300 NVL72, uses NVLink to achieve frontier intelligence, high reliability, and the lowest token cost, according to NVIDIA's congratulations. — via 1

2. AI Products and Adoption

1. Gemini app surpassed 1 billion monthly active users, becoming Google's 14th product to reach that milestone and its fastest-growing one. A celebratory event for builders, researchers, and contributors is planned for August 20 in San Francisco. — via 1 2
2. ChatGPT's Linux desktop app is now in preview, bringing ChatGPT, ChatGPT Work, and Codex to Linux with integrated project and browser workflows. — via 1
3. Transformers.js passed 10 million monthly downloads, becoming the most popular open-source library for running AI models in browsers. Local AI is increasingly favored for privacy and cost, especially amid compute shortages and cyberattack risks. — via 1
4. Runway's Agent added connectors for Figma, Dropbox, and Notion, unifying design and files for all paid plans. Separately, Seedance 2.5 is live on Runway, supporting 50 unique character references and generating up to 30-second music-synced clips. — via 1 2

3. AI Business and Funding

1. Lovable raised $400M at a $13.3B valuation, with annualized revenue exceeding $500M (up 7x YoY), 900M monthly visits, and two-thirds of Fortune 500 companies using the platform. Deedy, who led the round, argues the platform's real-revenue users and integrated Stripe/SEO capabilities make the "no moat" thesis wrong. — via 1 2
2. Ref, a shared space for teams to decide what agents to build before writing code, is now in open beta with $4M in funding. Its announcement introduces "Velocity Sickness" — teams feeling fast but lacking impact — and warns that letting agents make key decisions is the most dangerous failure mode. — via 1
3. Mistral AI says it is integrating reasoning infrastructure, open-source models, and long-term commitment to help Europe control its AI future and set a roadmap for the world. — via 1
4. Frontier classification model rd-signal-2 claims to be 1600x cheaper than GPT 5.6 Sol, with a free trial at @raindrop_ai. It offers custom classifier training/hosting APIs with zero data retention. — via 1 2

4. Research, Insights, and Trends

1. A paper exploiting an API vulnerability to extract hidden reasoning from frontier models is highlighted by swyx as one of the year's most important. Swyx notes the methodology is poorly explained and provides his own notes and distillation guidance. — via 1
2. Epoch AI's open math problem set advances: Levent Alpoge constructed 12 Hadamard matrices of order below 2000, the 4th solved among 50 open problems. — via 1
3. Ethan Mollick's research cautions against judging AI leadership by a single source: OpenRouter data appears to favor open-weight models, but submissions to Pangram overwhelmingly come from ChatGPT, with Claude's share growing. — via 1
4. Hamel Husain worries that AI's slow progress in long-form technical/scientific writing may hinder open-ended scientific research. He advocates inspecting systems directly — reading prompts and finding silly failures — rather than hoarding resources. — via 1 2
5. Deedy predicts a startup creation wave from big-lab departures in San Francisco, with massive capital flooding into "neolabs" and moonshot ideas. However, many founders will cluster in crowded domains like quantitative trading, drug discovery, materials, robotics, and hardware, leading to intense competition. — via 1