Startup Arcee AI has developed Trinity Large, an open-source Large Language Model (LLM) with 400 billion parameters.

Despite having a relatively small number of active parameters (17B), the model boasts a 512K token context window and aims for strong performance in agent harnesses, complex toolchains, and creative scenarios.

Trinity Large is offered in three checkpoints: 'Preview,' 'Base,' and 'TrueBase.' The 'TrueBase' version, in particular, serves as a true base model without alignment or instruction tuning, offering value to the research community.

The model's training took approximately 33 days, with an estimated cost of $20 million. While some evaluations suggest its performance may lag behind other state-of-the-art models, its unique architecture and open-source availability contribute to the diversity of AI research.

👉 Read the full article on Hacker News / Reddit

Arcee AI has released Trinity Large, an open-weight LLM with 400 billion parameters. The model emphasizes reliability in agent harnesses, handling complex toolchains, and excelling in creative scenarios.

Trinity Large features 13B active parameters and a 512K token context window. It is available in three versions: 'Preview,' 'Base,' and 'TrueBase,' each catering to different use cases.

The 'TrueBase' checkpoint is a pure base model without alignment or instruction tuning, designed for researchers to deeply explore what the model learned during its pre-training phase.

The training data includes over 8 trillion tokens of synthetic data across web, code, math, reasoning, and multilingual domains, utilizing state-of-the-art data curation approaches.

👉 Read the full article on Arcee AI / Reddit

OpenAI has launched the “OpenAI for Science” team, aiming to accelerate scientific research. This team focuses on leveraging LLMs like GPT-5 to assist scientists in finding connections to existing work, sketching mathematical proofs, and testing hypotheses.

Scientists report that GPT-5 is useful for brainstorming, discovering obscure references, and analyzing data faster, but mistakes and hallucinations remain concerns.

OpenAI is working on making models more epistemologically humble and implementing self-critique workflows. This initiative follows similar efforts by Google DeepMind, positioning OpenAI to compete in the scientific AI space.

OpenAI views the scientific community as a place to build the “next great scientific instrument,” combining AI models with research tools to extend human curiosity and drive discovery.

👉 Read the full article on MIT Technology Review / OpenAI

Amazon is reportedly in negotiations to invest up to $50 billion in OpenAI. This potential transaction follows a period of rapid valuation growth in the AI sector and signals a shift in cloud computing partnerships.

Amazon CEO Andy Jassy and OpenAI CEO Sam Altman are conducting direct discussions, which include integrating Amazon's hardware into OpenAI's infrastructure.

This investment is significant in the competitive landscape of cloud services and the semiconductor industry. While Amazon has historically backed Anthropic, a key competitor to OpenAI, these negotiations suggest a broader strategy to secure Amazon Web Services (AWS) as a foundational provider for multiple leading generative AI models.

OpenAI reached a valuation of $500 billion in October 2025 and is currently engaged in a $100 billion funding round. This round may include strategic participants like Microsoft and Nvidia, as well as a potential $30 billion contribution from SoftBank.

👉 Read the full article on TechCrunch / The Wall Street Journal