Open‑source 27B runs on a 3090

- Community developers report a 27‑billion‑parameter dense model running inference on a single NVIDIA 3090 GPU, suggesting frontier‑level capability on consumer hardware. - Running a 27B dense model on a 24GB 3090 implies lower‑cost access to capable LLMs for hobbyists and small labs, not just hyperscalers. - The social notes link these open models to agentic lab experiments and a broader shift away from exclusive billion‑dollar training runs. (x.com)