NVIDIA Nemotron Labs Dory1.0 4B

NVIDIA-Nemotron-Labs-Dory1.0-4B

Quick answer: NVIDIA-Nemotron-Labs-Dory1.0-4B is a 4B-parameter looped (latent-recurrent) language model from NVIDIA that applies a shared block of layers repeatedly (1-4 loops) to spend more compute per token, trading inference cost for accuracy from a single checkpoint with no retraining. It supports up to 256K context, was pre-trained natively in NVFP4, and is released under the OpenMDW-1.1 license.

At a Glance

Where Dory1.0-4B leads

• Elastic inference compute: one checkpoint serves multiple compute budgets by changing loop count (1/2/4)
• Memory-efficient looping: last-loop KV caching keeps cache size flat regardless of loop count
• Strong IFBench (74.29–79.0 across loop counts) relative to similarly sized peers

Where it lags

• Trails larger models (Gemma4-26B-A4B, Qwen3.6-35B-A3B) on GPQA Diamond and HLE even at 4 loops
• Requires unreleased/PR-stage vLLM and llama.cpp support at publication time (not yet in a stable release)

Best for: general-purpose reasoning on a small footprint where inference-time compute/accuracy tradeoff flexibility matters.

What Dory1.0-4B Is

Dory1.0-4B is NVIDIA's first public "looped" (latent-recurrent) LLM: the layer stack is split into Input, Recurrent, and Output groups, and the shared Recurrent block (20 layers) is applied repeatedly between 8-layer Input and Output blocks, with the input reinjected at every loop. This lets the model spend extra compute on harder tokens without adding parameters or requiring a different checkpoint per compute budget.

It was pre-trained on 6T tokens (first ~4.5T in NVFP4, remainder in BF16) plus 300B tokens of context extension to 256K, then went through supervised fine-tuning and multi-environment GRPO reinforcement learning (math, code, science, instruction following, tool use).

Specifications

Pricing

Open weights — self-hosting cost only; available in BF16, MXFP8, NVFP4, and Q5_K_M (GGUF) formats.

Context Window

Dory1.0-4B has a 256K-token context window, extended via a dedicated 300B-token training phase — enough for long documents or extended multi-turn agent sessions in a single request.

Public Benchmark Scores

Scores above are reported by NVIDIA and shown for context. Figures are measured under NVIDIA's own harness (NeMo Gym / NeMo Evaluator SDK) and are not Benchgen measurements. The 4-loop configuration is used for the headline numbers above; 1-loop and 2-loop results are lower (see model card for the full matrix).

Dory1.0-4B vs Alternatives

Dory1.0-4B trades parameter count for inference-time compute flexibility — it won't match a 26-35B MoE model on raw reasoning benchmarks, but a single 4B checkpoint can be dialed from cheap (1 loop) to more capable (4 loops) without redeploying.

How Dory1.0-4B Performs on Real Agent Tasks

Terminal-Bench 2.1 and SWE-bench Verified scores improve substantially with more loops (12.5→22.2 and 50.3→61.8 respectively from 1 to 4 loops), suggesting the looped architecture's extra compute budget is especially useful for agentic/tool-use tasks requiring longer reasoning chains, though it still trails much larger dense/MoE models on these same tasks.

Related Models & Benchmarks

• Compare: Qwen3.6-35B-A3B · Gemma 4 26B A4B
• Benchmarks: SWE-bench Verified · GPQA Diamond · SciCode

Frequently Asked Questions

Excluded vendor benchmarks (bucket C, documented here):

• Terminal-Bench 2.1 — no matching local entity (only Terminal-Bench 3/4/Hard/Science-0.1 exist, a methodology mismatch, not a straight successor).
• GDPval-AA-V2 — NVIDIA's reported Elo (815–987 range across loop counts) uses a different, smaller reference-model comparison pool (12 selected models) than convention elsewhere on this benchmark; excluded pending confirmation the harness/reference set matches our existing gdpval-aa-v2 entity, rather than assumed compatible.

Specs and scores sourced from NVIDIA's official HF model card; third-party benchmark scores attributed inline. Last updated 2026-10-23.