Allen AI releases open-source programming agent SERA: as low as $400 to adapt to private code repositories

Artificial intelligence research institution Allen AI recently announced the release of an open-source programming agent family — SERA. This series of models aims to lower the barrier for enterprises and developers to introduce AI programming capabilities into their private code repositories, with a minimum training cost of only $400.

Outstanding Performance

In the latest SWE-Bench-Test Verified programming benchmark test (64K context), the strongest model in this series, SERA-32B, successfully solved 54.2% of the problems. This achievement not only surpasses similar open-source models but is also on par with industry-leading closed-source models under certain conditions.

Very Low Training Threshold

According to data provided by Allen AI, SERA has extremely high training efficiency:

Very low cost: Only about 40 GPU days of training. It costs $400 to reach the level of mainstream open-source models, and $12,000 to achieve performance comparable to top commercial models.
Technological innovation: SERA adopts a simplified training method called **“Soft-verified Generation”**. This technology breaks the previous limitation that required “completely correct code examples,” making it more feasible to fine-tune on incomplete or private data.

Usability and Open Source Ecosystem

To facilitate quick integration for developers, Allen AI stated that the SERA model can seamlessly work with Claude Code, requiring only two lines of code to start.

Currently, all models, source code, and training instructions of SERA have been released on Hugging Face, and they follow the Apache 2.0 open-source license. This means that both small startups and large enterprises can freely build customized programming assistants tailored to their specific business logic based on this technology.

Valuation soars to $6.6 billion: AI programming star Lovable completes Series B funding, Google and NVIDIA join together

Swedish AI startup Lovable completes a $330 million Series B funding round, with a valuation of $6.6 billion, a more than threefold increase since July. The round was led by CapitalG and Menlo Ventures, with NVIDIA, Salesforce Ventures, and Deutsche Telekom participating. The company's core technology is text prompt generation AI.

Amazon appoints long-time AWS executive Peter DeSantos to lead new AI organization

Amazon CEO Andy Jassy announced that Peter DeSantos, a senior AWS executive, will lead a newly established artificial intelligence organization, focusing on AI models, chip development, and quantum computing to enhance AI tool performance. DeSantos has worked at Amazon for 27 years, including 8 years as Senior Vice President at AWS. AWS supports about one-third of global internet services, and this move aims to strengthen Amazon's competitiveness in the enterprise AI field.

DeepSeek releases V3.2-exp model, pioneering sparse attention mechanism significantly reduces AI inference costs

DeepSeek releases the experimental model V3.2-exp, which adopts an innovative 'sparse attention' mechanism to significantly reduce the cost of long context inference. The model is now available on Hugging Face and GitHub. The core is the 'lightning indexer' and optimized attention mechanisms to improve processing efficiency. This breakthrough technology is expected to promote the development of AI in the field of long text processing.

Google announces that Gemini custom assistant Gems is now available for sharing

Google announced on Thursday that its customized AI assistant Gemini Gems is now available for users to share. This feature was initially launched as part of a paid subscription service, allowing users to create AI chatbots tailored for specific tasks, such as learning coaches, writing editors, or programming partners. Google stated that users can now easily share their Gems with friends, family, or colleagues, just like sharing Google Drive files. This move aims to make these customized AI tools more accessible to a wider audience and avoid duplication.