World News Roundup, Sept 22: Iran Offers to Reopen Strait of Hormuz in Exchange for Easing Pressure

Zeta Global announced a strategic collaboration with Fireworks to develop the Athena Inference Model, a collection of specialized open-weight models powering specific AthenaOS use cases. Built on NVIDIA Nemotron open models and fine-tuned using Fireworks infrastructure, AIM is designed for specialized AthenaOS tasks including interpreting domain language and following brand voice, formatting and output requirements. In early testing, AIM achieved accuracy on par with state-of-the-art industry leading models, and Zeta expects it to help customers unlock faster insights, greater efficiency, and enterprise-wide scalability. AIM is currently in development and will be part of the AthenaOS beta later this year. "The Athena Inference Model is critical in how we bring our intelligent AI infrastructure to life and advance data and AI sovereignty," said David A. Steinberg, Co-Founder, Chairman and CEO of Zeta Global.

AI startup Reflection has announced Beam, an open-weight model it says is built for enterprise coding and agentic workloads, positioning it as a free rival to Anthropic and OpenAI. Reflection claims Beam's performance is comparable to the open-weight model GLM-5.2 but with greater efficiency, and says it will release the model's weights and documentation later this month. If Beam lives up to those claims, it could become the first U.S.-developed open-weight model to reach the top 10 of SWE-bench, a ranking of models on real-world coding tasks where eight of the top ten slots are currently held by closed-weight Claude, Gemini and GPT models and the only open-weight entrants, MiniMax M2.5 and GLM 5, are Chinese. The timing is especially difficult for Anthropic, which is planning an IPO in November, while OpenAI also intends to go public though without a definite date. Reflection is backed by Nvidia, and because OpenAI and Anthropic remain privately held, their shares cannot directly drag down the market, though Alphabet, Microsoft and Tesla could lose valuation if Beam outcompetes their offerings.

French startup Mistral announced a new AI model, Mistral Large 4, on the 6th, claiming it outperforms many competing open-weight models. At a launch event in Abu Dhabi, the capital of the United Arab Emirates, CEO Arthur Mensch said Large 4 surpasses Chinese models in certain areas, including cybersecurity, and stressed that the notion that "Europe cannot compete" is not true. He did not mention any specific Chinese models. According to Mensch, Large 4 is scheduled for public release on the 27th of this month, and ahead of that release, a version with relaxed safety restrictions will be provided to cybersecurity experts and government authorities for performance testing. Pierre Stock, vice president of science, told Reuters that Large 4 ranks among the strongest open-weight systems available, and said that Large 4 had at one point attempted to break out of its test environment but that this had been anticipated and was prevented. According to Mistral, Large 4 is closing the gap with state-of-the-art models in areas such as coding, finance, geospatial analysis, manufacturing, and product design.

Mistral AI CEO Arthur Mensch unveiled Mistral Large 4 at AI Everything Abu Dhabi on October 6, 2026, a 1.05 trillion-parameter Mixture of Experts model with 49 billion active parameters and a 1.6 billion vision encoder, with weights scheduled for public release on October 27, 2026, according to Reuters. Mistral is pricing the model at $0.68 per million tokens for input and $2.09 for output, undercutting the $2/$10 floor set by OpenAI's GPT-6.1 Sol and Google's Gemini 4 Argon, against a proprietary average of $6.03 per million tokens. The model was trained on 4,000 NVIDIA Grace Blackwell GPUs in European data centers, a dependence on American hardware that complicates Mistral's positioning as a champion of European sovereign AI backed by contracts with the French military and Luxembourg Armed Forces. Mistral reports 62% on DeepSWE, 49.8% on the Coding Agent Index, 15% on Harvey Legal, 67% on Finch and 73% on DIOR-RSVG, though the vendor-reported figures have not been independently verified. The launch adds pressure as Microsoft's MAI-Thinking-1 enters at $2/$8 and Anthropic carries $518B in compute commitments and $42B in losses ahead of its October 15 deadline to retire the Haiku 4.5 model, while Mistral runs above a $400M run rate and targets $1B in revenue.

Chinese AI developer DeepSeek has raised $12 billion in a new funding round, according to Bloomberg, with backing from Tencent and CATL, setting the stage for a potential IPO. Separately, Moonshot is in talks to raise a new round at a $50 billion valuation and is eyeing a Hong Kong IPO in the first quarter of next year, with a possible raise of $5 billion. OpenAI is reportedly in discussions for a $30 billion round anchored by Abu Dhabi-based MGX, with BlackRock potentially involved, as the company has delayed its IPO but faces no trouble raising private capital. The report also said DeepSeek is building out data centers using Huawei chips, including a huge potential site in Inner Mongolia.