Parameter Golf: openai's efficiency challenge. OpenAI has launched Parameter Golf, a global challenge to build the most efficient pretrained model under strict limits: 16 MB size and 10 minutes of training on 8xH100 GPUs. With compute credits available and job opportunities for standout participants, the contest blends technical innovation with recruitment potential. OpenAI has unveiled a new open research competition called Parameter Golf, aimed at pushing the boundaries of model efficiency. The challenge asks participants to minimise held-out loss on a fixed FineWeb dataset while adhering to strict limits: a combined 16 MB cap for weights and training code, and a 10-minute training budget on 8xH100 GPUs. This unusual set of constraints is designed to test ingenuity in model design and optimisation rather than brute computational force. Entrants are provided with a GitHub repository containing a baseline model, dataset, and evaluation scripts. They are expected to fork the repository, improve the model within the given size and compute caps, and submit their work via pull request. Submissions must include code, logs, scores, and a short write-up. Once approved, results are merged and automatically reflected on the leaderboard. To support experimentation, OpenAI has partnered with Runpod to offer compute credits. These range from quick-start credits worth around $25 (eight compute hours) to advanced competitor grants valued at $1,000 (320 compute hours). Credits are limited and subject to eligibility checks, but they provide a crucial resource for participants iterating on their approaches. Beyond the technical challenge, Parameter Golf serves as a talent-spotting exercise. OpenAI has stated that exceptional researchers and engineers who excel may be invited to interview for roles at the company. Winning approaches could also be showcased publicly, giving participants visibility in the AI research community. The competition is open to individuals aged 18 and above in supported jurisdictions. While OpenAI employees can participate, they are not eligible for compute credits. Participants can join discussions on OpenAI's official Discord, where a dedicated channel has been set up for collaboration and updates. This initiative reflects OpenAI's belief that innovation can come from anywhere. By framing efficiency as the ultimate test, the company is encouraging creative solutions that may influence future directions in model design. For aspiring researchers, Parameter Golf is not only a chance to demonstrate technical skill but also an opportunity to connect with one of the leading organisations in artificial intelligence.

Runpod, an AI cloud platform, has released its first State of AI Report, revealing significant shifts in production AI usage across 183 countries. The report, based on anonymised platform traffic and GPU utilisation data, shows Alibaba's Qwen has overtaken Meta's Llama as the most widely deployed self-hosted large language model. The data indicates a market focus on efficiency and modularity, with nearly 70% of image workflows running through ComfyUI and video upscaling workloads outpacing raw generation by a 2:1 ratio. Nvidia B200 usage scaled 25-fold in 2025, with supply projected to nearly quadruple by mid-2026. The report also reveals that whilst the US leads in user base, India has become the second-largest market. vLLM has emerged as the standard for LLM serving, powering 40% of all LLM endpoints.