Open-Source AI

Today's Stories

1. 1

   DeepSeek's open-source reasoning model has reached the #2 rank among open-weights models, challenging the dominance of larger commercial systems.

   DeepSeek released an open-source reasoning model that now ranks as the #2 open-weights reasoning model. The model operates on 27% of FLOPs compared with DeepSeek-V3.2 and supports 1M tokens context length, up from 128K in V3.2. Open-source reasoning models that match or approach closed commercial systems in performance could shift where companies and developers turn for AI reasoning tasks—the step where an AI produces an answer to complex questions. A high-ranking open model reduces dependence on proprietary offerings.

   The model's efficiency (27% of FLOPs compared with DeepSeek-V3.2) and extended context window (1M tokens) are the capabilities to monitor as other organizations benchmark it against existing leaders in the reasoning space.

2. 2

   Krea releases open-weights AI image generation models to address complaints that AI-generated visuals look repetitive and unoriginal.

   AI creative tools startup Krea has opened the weights to two new image generation models, Krea 2 Raw and Krea 2 Turbo, under a custom license. Both models are available for download on Hugging Face. Firms with more than 50 seats must pay for Enterprise usage, and all users must implement technical safeguards to prevent generation of illegal materials, non-consensual intimate imagery, child sexual abuse material, or defamatory assets. Enterprises are already integrating AI-generated images into production workflows, but there is growing commentary that AI imagery looks monotonous and unoriginal—what users call "AI slop." Krea's move to open the model weights suggests the company believes transparency and broader access can help address these quality and originality concerns.

   Both Krea 2 Raw and Krea 2 Turbo are available now on Hugging Face. The company states the models provide more visual variety than typical alternatives, though the body does not specify performance benchmarks or pricing details for the Enterprise tier.

3. 3

   Papers with Code, the research discovery platform, is reviving features from its original version—including achievement badges and a new trending score—to help researchers surface and build on each other's work.

   Papers with Code has added support for SOTA (state-of-the-art) badges that appear on papers scoring in the top 3 of a given benchmark, and is introducing a new trending score metric. The platform is rolling out these features as part of a broader revival effort by Hugging Face's open-source team. The platform's creators describe this as a response to what they call a return to "the age of research," where researchers need to discover each other's work and build collaboratively. These features make it easier to identify which papers and models are performing best, helping the research community stay informed about the latest advances.

   The SOTA badges are visible across all paper feeds on the site—for example, the video-classification task page—making top-performing research discoverable alongside individual papers.

4. 4

   OpenAI launches 'Patch the Planet' initiative with security firm Trail of Bits to help open-source projects identify and fix vulnerabilities using AI tools.

   OpenAI announced a program called 'Patch the Planet' in partnership with security company Trail of Bits. Security engineers from Trail of Bits will work directly with open-source maintainers to review code issues, develop patches and tests, and build reusable workflows—supported by OpenAI's security tools like Codex Security. Open-source software underpins the commercial software industry, but many projects suffer from poor security due to their decentralized structure. Bugs in widely used open-source utilities can create major problems for businesses; the article cites the log4j vulnerability as an example. This initiative aims to reduce the burden on maintainers who already struggle to handle security reports with limited resources.

   The program's design prioritizes reducing maintainer burden by having security engineers review findings before they reach project teams, rather than adding to their workload. It remains unclear how the initiative will scale and function long-term, though the effort appears aimed at helping the open-source community better protect itself against automated vulnerability discovery and exploitation.

5. 5

   Reflection AI signs $150 million（約240億円） monthly computing deal with SpaceX through 2029, securing access to advanced AI chips for model training.

   Reflection AI, an open-source AI startup, has signed an agreement with SpaceX to access computing capacity at SpaceX's Colossus 2 data center. The startup will gain immediate access to Nvidia GB300s (AI chips used to train and run advanced models) and will pay SpaceX $150 million（約240億円） per month beginning July 1, 2026, through 2029. Large-scale computing power is a critical bottleneck for AI development. By securing dedicated access to specialized chips and data center capacity at a fixed cost, Reflection AI can plan its model development with certainty, rather than competing for limited resources in a constrained market.

   The monthly payment of $150 million（約240億円） beginning July 1, 2026, and the agreement's duration through 2029 set a clear financial and timeline commitment. Neither SpaceX nor Reflection AI has responded to requests for comment on the deal.

6. 6

   An open-source semantic caching proxy called AI-Gateway has been shared on GitHub, aiming to reduce LLM API costs by caching 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 reduce costs associated with large language model API usage by caching semantically similar requests. LLM API calls can be expensive for businesses and developers, and semantic caching (storing answers to similar queries rather than reprocessing them) offers a way to lower those expenses without sacrificing response quality. An open-source approach means teams can deploy and modify it themselves rather than relying on a third-party vendor.

   The project is available now on GitHub for developers and teams interested in testing or deploying semantic caching to their LLM workflows. Adoption and community feedback will determine whether this approach becomes a standard cost-reduction practice in the industry.