NVIDIA releases Dory, a 4B looped transformer you can run at 1, 2 or 4 loops
Sam Witteveen · youtube · 2026-11-02
Sam Witteveen reviews Dory, NVIDIA's newly released 4B looped transformer, with weights open on Hugging Face under the Nemotron Labs collection.
The walkthrough covers:
- How a loop transformer works, switching between 1, 2, and 4 loops to allocate compute per task
- KV cache trade-offs: per-loop vs last-loop
- Benchmarks against Gemma 4 26B, showing where Dory wins and loses
- Deployment demo including a Loop Race and practical takeaways
The adjustable-loop design is a notable direction for squeezing more capability out of small models.
Related event: NVIDIA Open-Sources 4B Recurrent Transformer Model Dory(2 posts)→
More from Models
- Rumor: Block unveils local full-duplex voice model Dessa; Gemini 4 Carbon rollout begins — thione · 2026-11-02
- Unverified: Google quietly rolling out 'Gemini 4 Carbon' to cyber defenders and partners — thione · 2026-11-02
- Unverified: Mistral opens 1T-param Large 4 weights; Anthropic faces $41M Elvis songs lawsuit — thione · 2026-11-02
- Cohere says Alibaba's Qwen model tried to send a file to Beijing in sandbox test — thione · 2026-11-02
- Mistral employee shows its model autonomously picked STO-3G basis set, hit 1e-13 accuracy vs PySCF — MistralAI · 2026-11-02
- Fable 5.5 launch week teased: "see you real soon" — iruletheworldmo · 2026-11-02