Open-Source AI

Today's Stories

1. 1

   Reflection AI signs $150 million（約240億円） per month computing deal with SpaceX to access advanced chips through 2029.

   Reflection AI, an open-source AI startup, has agreed to a deal with SpaceX granting it access to Nvidia GB300 chips at SpaceX's Colossus 2 data center. The startup will pay SpaceX $150 million（約240億円） per month beginning July 1, 2026, through 2029. AI startups need substantial computing capacity to train and run advanced models. This deal gives Reflection guaranteed access to high-end hardware at a major data center, removing a potential bottleneck for the company's development.

   The agreement begins July 1, 2026, and runs through 2029, locking in computing costs for the startup over a multi-year period.

2. 2

   An open-source semantic caching proxy called AI-Gateway has been shared on GitHub with the goal of reducing LLM API costs by caching and reusing similar requests.

   A developer has released AI-Gateway, an open-source semantic caching proxy available on GitHub (https://github.com/Arnab758/ai-gateway). The tool is designed to intercept and cache API requests to large language models in a way that recognizes semantic similarity, allowing duplicate or similar requests to be served from the cache instead of sending them to the API. Large language model API calls can be expensive, and many businesses making repeated queries face high costs. By caching semantically similar requests, AI-Gateway can reduce the number of API calls needed, which in turn lowers spending on LLM services. This addresses a practical pain point for teams relying heavily on AI APIs.

   The project is hosted on GitHub and appears to be in early stages, with minimal engagement so far (0 comments and 1 point at the time of posting). Whether the semantic caching approach proves effective in real-world use cases, and how the open-source community adopts or improves the tool, will determine its practical impact.

3. 3

   OpenAI launches Patch the Planet, a free security program to help open-source projects fix vulnerabilities faster, as AI bug-hunting tools overwhelm volunteer maintainers.

   OpenAI, in partnership with security firm Trail of Bits and vulnerability management firms HackerOne and Calif, announced Patch the Planet on Monday. The program offers free security consulting to open-source maintainers, helping them find and patch vulnerabilities and incorporate AI security tools into their development process. More than 30 open-source projects are already participating, and the program has uncovered hundreds of bugs and produced dozens of patches in its first week. Open-source developers—typically volunteers with limited resources—are struggling to keep up with bug reports. The rise of AI vulnerability hunting has made this backlog feel unmanageable as AI-generated reports pile up, pulling attention away from critical flaws. Patch the Planet aims to reduce this burden by subsidizing token costs (OpenAI has been subsidizing Codex Security scanner usage "to the tune of 20 trillion tokens") and providing human expertise to offset maintainers' workload.

   Participants receive six months of free ChatGPT Pro and six months of Codex Security access, plus infrastructure improvements they can use long-term. Trail of Bits, funded by OpenAI with unmetered model access, plans to continue the program long-term. The announcement comes as OpenAI competes with Anthropic, which had to pull its Mythos 5 and Fable 5 models from the market earlier this month due to export controls over cybersecurity capabilities.

4. 4

   Open source AI startup Reflection AI signs $6.3 billion（約1兆円） compute deal with SpaceX, betting that publicly available AI models will gain ground as governments restrict proprietary alternatives.

   Reflection AI will pay $150 million（約240億円） per month beginning July 1, 2026 through 2029 for access to Nvidia's GB300 AI chips at SpaceX's Colossus 2 data center near Memphis, Tennessee. The deal is worth up to $6.3 billion（約1兆円） and either company can end the contract with 90 days' notice after the first three months. Reflection is positioning open-weight AI models—which publicly release their trained parameters—as a viable alternative to closed proprietary systems. The company cited the U.S. government's recent ban of Anthropic's closed models as evidence that nations and enterprises are recognizing risks in depending exclusively on proprietary AI, making open source infrastructure increasingly strategic.

   The deal is smaller than SpaceX's agreements with Anthropic ($1.25 billion（約2000億円） per month) and Google ($920 million（約1500億円） per month), both also running through 2029. Reflection, founded in 2024 by two former Google DeepMind researchers, describes this as one of the largest announced open AI infrastructure commitments to date.

5. 5

   OpenAI launches Patch the Planet, a program to help open-source maintainers identify and fix software vulnerabilities using AI assistance and expert review.

   OpenAI has introduced Patch the Planet as part of its Daybreak initiative. The program is designed to help open-source software maintainers find, validate, and fix vulnerabilities with AI assistance and expert review. Open-source software is widely used across technology stacks, and unpatched vulnerabilities can create security risks. By providing maintainers with AI tools and expert support, the program aims to reduce the time and resources required to address security issues, potentially strengthening the security of the broader software ecosystem.

   The program is part of OpenAI's Daybreak initiative, signaling the company's broader commitment to supporting open-source maintainers. Participation and adoption rates will indicate how widely the program reaches the maintainer community and whether it measurably improves vulnerability remediation timelines.

6. 6

   PaddleOCR releases PP-OCRv6, a family of lightweight text-recognition models spanning 1.5M to 34.5M parameters with support for 50 languages, improving detection and recognition accuracy over its predecessor while remaining deployable across multiple runtime environments.

   PaddleOCR launched PP-OCRv6, a model family with three tiers (tiny, small, medium) ranging from 1.5M to 34.5M parameters. The medium tier achieves 86.2% detection Hmean and 83.2% recognition accuracy, improving text detection by +4.6 percentage points and text recognition by +5.1 percentage points compared with PP-OCRv5_server. The medium and small tiers support 50 languages, including Simplified Chinese, Traditional Chinese, English, Japanese, and 46 Latin-script languages. Businesses handling documents, screenshots, industrial labels, and multilingual content need accurate text extraction at scale. PP-OCRv6 offers a production-ready alternative to larger models, with the flexibility to choose model size based on deployment constraints—edge devices can use the tiny tier, while accuracy-critical server pipelines can use the medium tier. The unified multilingual support in a single model family reduces the complexity of managing separate OCR systems per language.

   PP-OCRv6 is available now on the Hugging Face Hub in multiple formats (safetensors, Paddle inference models, and ONNX models) and can be evaluated via the online demo at PP-OCRv6 Online Demo. It supports three inference backends—PaddleOCR's native Paddle Inference, Transformers (for PyTorch workflows), and ONNX Runtime—enabling deployment across different runtime environments without retraining.