RWKV-7 G1g is here: the world's best pure RNN LLM, and a competitive LLM in general. Try for bsz16 7B inference. G1h in June p.s. const 15000+tp…

RWKV-7 G1g (7B) is released, a pure RNN architecture claiming state-of-the-art among RNN-based LLMs and competitive with transformer models generally. Achieves 15,000+ tokens/second decoding on a single RTX 5090 at batch size 16, via the Albatross inference engine on GitHub. Relevant for edge/long-context deployments where constant memory scaling versus sequence length matters over transformer alternatives.