NVIDIA Releases Nemotron Labs Dory 1.0 Open-Source Looped Language Model

Performance & Benchmarks

Despite its compact 4B parameter size, Dory 1.0 punches significantly above its weight class, particularly in reasoning-heavy tasks. In preliminary testing, the model demonstrates efficiency levels that rival much larger 7B and 13B parameter static models. Because the depth is adjustable, performance scales with the number of recurrent loops applied during inference.

On the MMLU (Massive Multitask Language Understanding) benchmark, Dory 1.0 achieves a score of 68.4% at its standard loop setting, outperforming several contemporary 7B models. In coding-specific evaluations, it hits 54.2% on HumanEval, making it a viable candidate for local code completion and assistance. For complex reasoning, its SWE-bench performance shows high potential for agentic tool-use when the loop count is maximized.

The evaluation methodology focused on comparing the 'elastic' performance against fixed-depth baselines. Benchmarks were conducted using a standardized NVIDIA H100 environment, ensuring that the throughput gains from the sliding-window attention were accurately captured relative to standard causal LLMs.

• MMLU: 68.4% (standard loop)
• HumanEval: 54.2%
• High efficiency in reasoning-to-parameter ratio
• Evaluated on NVIDIA H100 clusters

API Pricing

As Dory 1.0 is an open-source release intended for local deployment and self-hosting, there is no centralized NVIDIA-hosted API pricing for this specific model. Developers are encouraged to deploy the model on their own infrastructure using optimized runtimes.

The model is shipped in multiple formats to facilitate various deployment scenarios, including BF16, MXFP8, NVFP4, and GGUF. This allows for highly cost-effective deployment on consumer hardware or enterprise-grade clusters depending on the quantization level chosen.

• Direct API Pricing: N/A (Open Source)
• Deployment: Self-hosted via vLLM or llama.cpp
• Quantization Support: BF16, MXFP8, NVFP4, GGUF

Use Cases

The primary strength of Dory 1.0 lies in its adaptability. The 'elastic depth' makes it uniquely suited for edge computing and mobile applications where compute resources may fluctuate. An agentic system could use a low-loop setting for simple intent classification and automatically scale to a high-loop setting for complex logical reasoning.

It is also an ideal candidate for RAG (Retrieval-Augmented Generation) pipelines where low-latency response is critical. The hybrid attention mechanism allows it to process retrieved context windows efficiently. Furthermore, its small footprint makes it a perfect fit for local coding assistants and real-time chat interfaces that require high privacy and low hardware requirements.

• Edge AI and Mobile Deployment
• Dynamic Agentic Workflows (Variable Reasoning)
• Low-latency RAG systems
• Local Coding Assistants

Getting Started

To get started with Dory 1.0, developers should head to Hugging Face to download the weights. Note that because of the novel looped architecture, you cannot use a standard LlamaForCausalLM runtime. You will need to use either vLLM or llama.cpp, specifically utilizing the dedicated 'looped-layers' pull request to support the recurrent layer application.

For those looking to experiment with the most efficient quantization, the NVFP4 versions are highly recommended for NVIDIA Blackwell or Hopper architectures, while GGUF remains the gold standard for CPU/Apple Silicon deployment via llama.cpp.

• Model Access: Hugging Face
• Required Runtime: vLLM or llama.cpp (with looped-layers PR)
• License: OpenMDW-1.1 (Commercial & Non-Commercial)

Sources