NVIDIA Releases Dory1.0-4B: A Recurrent 4B LLM with Dynamic Loops

NVIDIA Releases Dory1.0-4B: A Recurrent 4B LLM with Dynamic Loops

At a Glance

Item Value
Repository nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16
Family guide NVIDIA-Nemotron-Labs-Dory1.0 guide (2 articles)
Publisher guide NVIDIA: models and licenses
Published 2026-10-17
License openmdw-1.1
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

NVIDIA has released “Dory1.0-4B", a looped (latent recurrent) large language model. This model adopts a normalized-GPT architecture that applies a shared recurrent block multiple times between input and output blocks. It features elastic inference compute that allows the number of layer loops to be changed dynamically during inference, and supports a context length of up to 256K tokens.

Specifications

  • Parameters: 4B
  • Architecture: Looped (latent recurrent) normalized-GPT with shared recurrent blocks, hybrid structure of sliding window attention (SWA) and full attention layers
  • Context length: Up to 256K tokens

Performance

The comparison table with other models from the model card is shown below. Qwen3.5-4B, Qwen3.5-9B, Gemma 4 26B A4B, and Qwen3.6-35B-A3B are listed as comparison targets.

→ Scroll horizontally to see all columns

Task Dory1.0-4B / 1 loop Dory1.0-4B / 2 loops Dory1.0-4B / 4 loops Qwen3.5-4B Qwen3.5-9B Gemma 4 26B A4B Qwen3.6-35B-A3B
Coding & Agentic
SciCode 27.3 34.94 35.11 25.82 28.21 40.32 37.24
SWE-bench Verified 50.3 59.5 61.8 54.7 61 57.40 73.4†
Terminal-Bench 2.1 12.5 18.1 22.2 15 24 35 43
τ³-bench (Banking) 11.6 12.89 14.69 5.03 7.22 13.66 9.54
GDPval-AA-V2 (Elo) 815 946 967 424 546 751.6 987
Reasoning
GPQA Diamond (no tools) 69.19 72.41 72.16 77.71 80.93 80.43 85.35
HLE (text-only, no tools) 8.25 10.89 10.43 11.31 15.43 17.79 19.18
Instruction Following
IFBench 74.29 77.94 79 53.83 59 77.89 57.44
Long Context
AA-LCR 63 71 73 64.7 68.4 64.6 68.5
General Knowledge
AA-Omniscience (accuracy) 10.03 10.08 10.03 15.02 18.05 21.13 19.95

According to measurements by the publishers, accuracy tends to improve as the number of loops is increased. Particularly in coding and agentic tasks (such as SWE-bench Verified), the score increases significantly from 50.3 to 61.8 in the 4-loop setting, outperforming the comparable Qwen3.5-4B in some items. On the other hand, in difficult scientific challenges like GPQA Diamond and HLE measuring extremely difficult questions, the figures are lower compared to larger models such as Qwen3.5-9B and Gemma 4 26B A4B, falling short of large-scale models in areas of reasoning ability and general knowledge. Regarding performance changes by the number of loops, the improvement margin is large at the 2nd loop for most tasks, and it has the characteristic of being further optimized by increasing to 4 loops.

Strengths and Use Cases

Dory1.0-4B adopts a looped architecture with a shared recurrent block sandwiched between input and output blocks, and its biggest feature is that it can adjust the trade-off between accuracy and computational cost by dynamically changing the number of loops during inference. In model cards and other materials, it has undergone multi-environment reinforcement learning (GRPO) across diverse settings such as coding, mathematics, science, instruction-following tasks, multi-step tool use, multi-turn conversations, and structured outputs, giving it strengths in agentic tasks and instruction-following capability. It supports English and handles long-context processing up to 256K tokens.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 4.0B parameters

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) Q6_K 3.0GB 3.7GB
8GB (RTX 4060 / 3060 Ti, etc.) Q8_0 3.9GB 4.7GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: not registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. File sizes are measured from the converted build DevQuasar/nvidia.NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16-GGUF. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but DevQuasar/nvidia.NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16-GGUF provides a GGUF build you can use.

License — openmdw-1.1: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compression: the Q6_K build measures 6.48 bits per weight — about 40% the size of the original 16-bit weights, calculated by this site from the actual file sizes.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (openmdw-1.1). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Recent Models in the Same Size Class

Models with up to 4B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
openbmb/MiniCPM-V-4.7-35B-A3B-DSpark-gguf 389M 4GB apache-2.0 MiniCPM-V-4.7-35B-A3B-DSpark-gguf Draft Model for Speedup: 4GB+ VRAM (2026-11-17)
openbmb/MiniCPM-V-4.7-1B-DSpark-gguf 214M 4GB apache-2.0 MiniCPM-V-4.7-1B-DSpark-gguf Text Generation Model: 4GB+ VRAM (2026-11-17)
IFM/Jais-3-0.9B-GGUF 1.1B 4GB apache-2.0 Jais-3-0.9B-GGUF Bilingual Dense Model: 4GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-DSpark 328M 4GB lfm1.0 LFM2.5-24B-A2B-DSpark Speculative Decoding Draft Model: 4GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-DSpark-GGUF 328M 4GB lfm1.0 LFM2.5-24B-A2B-DSpark-GGUF Text Generation Model: 4GB+ VRAM (2026-11-17)

How to Get It

It is available on Hugging Face in multiple formats including BF16, MXFP8, NVFP4, and Q5_K_M formats. It is available for commercial and non-commercial use, complying with the OpenMDW License Agreement, version 1.1. No special license agreement (gated) is set for acquisition.

It can be run locally using inference engines such as vLLM and llama.cpp. When running with vLLM, start the server using pull request-based compatible versions or similar.

vllm serve --model nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 \
    --hf-overrides '{"n_recurrent_loops": 2, "recurrent_kv_cache_mode": "per_loop"}' \
    --max-model-len 262144 \
    --max-num-seqs 128 \
    --enable-prefix-caching \
    --async-scheduling \
    --reasoning-parser nemotron_v3 \
    --tool-call-parser qwen3_coder \
    --enable-auto-tool-choice

Alternatively, when using llama.cpp, download a GGUF-formatted quantized file (such as dory-q5_k_m.gguf), build from the corresponding source, and start the server.

./build/bin/llama-server \
    -m./dory-q5_k_m.gguf \
    --alias dory \
    --host 127.0.0.1 --port 8080 \
    -ngl 999 \
    --ctx-size 4096 --parallel 1 \
    --batch-size 512 --ubatch-size 512 \
    --cache-type-k f16 --cache-type-v f16 \
    --flash-attn off \
    --no-context-shift \
    --jinja --reasoning on \
    --temp 1.0 --top-p 0.95 --top-k 0 --min-p 0 --repeat-penalty 1 \
    --override-kv "dory.num_loops=int:2,dory.recurrent_kv_cache_mode=str:per_loop"

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-17 DevQuasar GGUF DevQuasar/nvidia.NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16-GGUF Q6_K 3.7GB (fits in 4GB VRAM)
2026-11-17 mlx-community MLX mlx-community/NVIDIA-Nemotron-Labs-Dory1.0-4B-MLX-8bit MLX 8bit 4.7GB (fits in 8GB VRAM)

File sizes of each build:

  • Available builds in DevQuasar/nvidia.NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16-GGUF: Q2_K 1.5GB / Q3_K_M 1.9GB / Q4_K_M 2.3GB / Q5_K_M 2.6GB / Q6_K 3.0GB / Q8_0 3.9GB
  • Available builds in mlx-community/NVIDIA-Nemotron-Labs-Dory1.0-4B-MLX-8bit: MLX 8bit 3.9GB

In addition, 10 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-17: Added converted builds to “Quantized and Converted Variants”: DevQuasar/nvidia.NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16-GGUF, mlx-community/NVIDIA-Nemotron-Labs-Dory1.0-4B-MLX-8bit
  • 2026-11-17: Updated the hardware requirements table with the actual file sizes of DevQuasar/nvidia.NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16-GGUF.