Ramanujan-1 — completely without Nvidia

The startup Essential AI, founded by Ashish Vaswani, has released its first model. This 8B model was trained on a mix of AMD MI300x and Google TPU v5p, with the entire training process running on Jax. The team boasts an MFU of 50% on MI300x, although it seems to be referring to bf16. In pretrain and instruct benchmarks, the model shows results comparable to Qwen 3 8B, despite using four times less compute for its training. However, it truly shines in agent coding — on the SWE-bench Verified, Ramanujan-1 achieves phenomenal results of 20.8 for its size. Other models in this weight category yield near-zero results, so they are usually not tested on such a complex benchmark. The startup team consists of about 20 people, so the release is very focused — the model was released in only one size, on a stock Gemma architecture, without a reasoning variant and with a tiny context of 32k. But a solid foundation for future releases has been laid, and we look forward to seeing the team scale up.