NVIDIA Nemotron Dory 4B model shows strong coding performance

NVIDIA has released its new 4B parameter model, Nemotron Dory 4B, which demonstrates strong performance on coding and agentic tasks, outperforming larger models like Qwen3.5-9B on the SWE-bench benchmark. The model utilizes a unique looped recurrent block architecture, allowing it to process tokens through multiple layers, which significantly enhances its capabilities for complex reasoning and instruction following. Despite its impressive performance, support for Dory 4B is not yet widely available in popular LLM frameworks like llama.cpp, vLLM, and Transformers, though a PR build for llama.cpp is available. AI

IMPACT Sets a new benchmark for small models in coding tasks, potentially influencing future small model development and agentic AI.

RANK_REASON Frontier-lab model release with system card and benchmark data. [lever_c_demoted from frontier_release: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• Hugging Face
• llama.cpp
• NeMo Evaluator
• NeMo Gym
• Nemotron Dory 4B
• NVIDIA
• NVIDIA Nemotron Labs Dory 1.0 4B
• OpenMDW License v1.1
• Qwen3.5-4B
• Qwen3.5-9B
• SWE-bench
• Transformers
• vLLM

• model release
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →