Reddit debates RWKV: faster and cheaper, but would you build an LLM on RNNs today?
Haghiri75 · reddit · 2026-08-15
A Reddit thread weighs whether RNNs — specifically the RWKV approach — still make sense for language generation. The poster's motivation is cutting the cost of LLMs on repetitive workloads like coding; their reading of the RWKV paper is that it grafts a QKV-style matrix mechanism onto a traditional RNN.
In personal tests, RWKV models ran faster on both Colab and gaming PCs, and stayed faster even when quantized and run on CPU via ollama. The question posed to the community: if you were training an LLM from scratch today, would you take this route?
More from Research
- Proposal: Buy Research Data Directly for Pretraining in Exchange for Compute — teortaxesTex · 2026-08-15
- SuperFlex Introduces Bending and Tapering for Improved Superquadric Point Cloud Decomposition (ECCV 2026) — CSProfKGD · 2026-08-15
- FLIM imaging reveals long-distance non-neural bioelectric patterns — drmichaellevin · 2026-08-15
- Open source closes the gap with closed labs: Quality gap now just months — togethercompute · 2026-08-15
- Maglev: Sliding Recurrent Memory improves long-context efficiency — UTEXAS · 2026-08-15
- AI-assisted research trends spark a surge in NeurIPS submissions and quality advances — PTenigma · 2026-08-15