NVIDIA Releases Dory 1.0 4B: Looped Latent-Recurrent Model with Elastic Inference Compute

The Three-Question Framework

What Happened?

NVIDIA Nemotron Labs published Dory 1.0 4B, an open-weight looped latent-recurrent language model that divides its architecture into 8 input layers, 20 shared recurrent layers, and 8 output layers under the OpenMDW-1.1 license.

Why It Matters

Dory enables elastic inference compute: by simply adjusting the number of times the shared 20-layer recurrent block is executed (from 1 to 4 loops), SWE-bench Verified jumps from 50.3% to 61.8% with zero model swapping or KV-cache memory expansion.

What Could Happen Next

Developers and edge deployment teams are porting Dory's last-loop KV caching routines into TensorRT-LLM and vLLM runtimes to evaluate latency tradeoffs in constrained workstation environments.

Executive Key Takeaways

Dory 1.0 4B employs a looped normalized-GPT (nGPT) architecture featuring 8 input layers, 20 shared recurrent layers, and 8 output layers.

SWE-bench Verified scores scale elastically with compute: 50.3% at 1 loop, 59.5% at 2 loops, and 61.8% at 4 loops, outperforming Qwen3.5-4B (54.7%) and Gemma 4 26B A4B (57.4%).

SciCode scores advance from 27.3% (1 loop) to 34.94% (2 loops) and 35.11% (4 loops), demonstrating latent-space reasoning without generating verbose chain-of-thought tokens.

Last-loop KV caching ensures memory usage remains constant regardless of the loop count, eliminating the quadratic KV cache penalty of conventional recurrence.

Natively pre-trained in NVIDIA's NVFP4 4-bit floating-point format and completed in BF16, Dory is available in BF16, MXFP8, NVFP4, and GGUF checkpoints.

NVIDIA Nemotron Labs announced the open-weight release of NVIDIA-Nemotron-Labs-Dory1.0-4B on October 16, 2026, introducing a looped latent-recurrent architecture designed to provide elastic inference compute within a compact 4-billion-parameter memory footprint. Published under the OpenMDW-1.1 license, the model is available across BF16, MXFP8, NVFP4, and Q5_K_M GGUF formats on Hugging Face.

Dory departs from conventional fixed-depth causal transformers by segmenting its layer stack into three functional groups: an 8-layer Input block, a 20-layer shared Recurrent block, and an 8-layer Output block. During inference, tokens pass through the input block, cycle through the shared recurrent block between one and four times, and finally exit through the output block. The original input embeddings are reinjected into each recurrent loop, enabling the model to refine latent representations without generating additional output tokens.

This architectural mechanism provides elastic inference compute from a single checkpoint. In empirical evaluations, increasing the loop count from one to four boosted SWE-bench Verified solve rates from 50.3% to 59.5% at two loops and 61.8% at four loops. At maximum loop depth, the 4B model surpasses larger fixed-depth baselines, including Qwen3.5-4B (54.7%) and the 26B Gemma 4 26B A4B MoE (57.4%). On SciCode, scores climb from 27.3% to 35.11%, while Terminal-Bench 2.1 improves from 12.5% to 22.2%.

A critical engineering barrier in looped transformers has historically been memory bloat: storing independent Key-Value (KV) cache entries for each loop iteration multiplies memory requirements. NVIDIA resolved this by implementing last-loop KV caching, which preserves only a single cache state per recurrent attention layer. As a result, running Dory across four loops introduces zero KV cache memory overhead relative to a single-loop pass.

The model was pre-trained natively in NVIDIA's NVFP4 4-bit floating point format on a normalized-GPT backbone before undergoing BF16 precision continuation and post-training on Nemotron post-training datasets. With support for a 256K token context window, Dory provides a practical template for resource-constrained robotics and edge coding agents where inference compute must scale dynamically according to query difficulty.

The Three-Question Framework

What Happened?

Why It Matters

What Could Happen Next

Executive Key Takeaways

Primary Sources & Citations