Nemotron Labs Dory 1.0 4B review, pricing and limits

Dory 1.0 4B ships in four formats, from a 7.95 GB full-precision file to a 3.20 GB NVFP4 file, so the weights alone fit on a single consumer GPU. It suits developers who want a small self-hosted English coding or agent model, and not teams that need images, other languages or a hosted API.

Nemotron Labs Dory 1.0 4B is NVIDIA's open-weight looped language model, scoring 59.5% on SWE-bench Verified at its recommended two loops, as measured by NVIDIA on the full-precision checkpoint. One 4B checkpoint reapplies a shared block of layers one, two or four times per token, so users trade accuracy against compute at launch time.

Provider: NVIDIA · Family: Nemotron Labs

More about NVIDIA on HokAI

Context window: 262,144 tokens

Input modalities: text, code, tool-calls · Output: text, tool-calls

About Nemotron Labs Dory 1.0 4B

Nemotron Labs Dory 1.0 4B is a looped language model that NVIDIA published on 16 October 2026 under its Nemotron Labs banner, with weights under the OpenMDW-1.1 licence. The release is four Hugging Face repositories (BF16, MXFP8, NVFP4 and a Q5_K_M GGUF); HokAI found no launch blog post, paper or hosted endpoint for it as of 20 October 2026. The BF16 file holds 3,975,229,440 parameters in a normalized-GPT stack. Its layers split into an 8-layer input block, a 20-layer shared recurrent block and an 8-layer output block, counting an attention layer and its MLP as one layer. The recurrent block is applied several times per token with the input fed back in on every pass, so the model spends more computation per token without adding weights. NVIDIA pre-trained it on 6 trillion tokens, spent another 300 billion tokens extending the context, then ran supervised fine-tuning and GRPO reinforcement learning across math, code, science, instruction following, multi-step tool use, multi-turn chat and structured output.

NVIDIA's model card reports results at 1, 2 and 4 loops from the same checkpoint, measured by NVIDIA on the BF16 weights with its NeMo Evaluator tooling. SWE-bench Verified climbs from 50.3 at one loop to 59.5 at two and 61.8 at four. Terminal-Bench 2.1 goes 12.5, 18.1, 22.2; IFBench goes 74.29, 77.94, 79; AA-LCR, a long-context test, goes 63, 71, 73. Not every row keeps rising: GPQA Diamond peaks at two loops (72.41 against 72.16 at four) and so does Humanity's Last Exam (10.89 against 10.43), which fits the card making two loops the default. Against Qwen3.5-4B in the same table, two loops is ahead on SWE-bench Verified (59.5 against 54.7), Terminal-Bench 2.1 (18.1 against 15) and IFBench (77.94 against 53.83), and behind on GPQA Diamond (72.41 against 77.71) and AA-Omniscience, a knowledge-recall test (10.08 against 15.02). At four loops its SWE-bench Verified score edges past the 61 listed for Qwen3.5-9B, though the 9B model still leads Terminal-Bench 2.1 (24 against 22.2) and GPQA Diamond (80.93). Gemma 4 26B A4B leads Dory on SciCode, Terminal-Bench 2.1, GPQA Diamond and Humanity's Last Exam but trails the two-loop model on SWE-bench Verified (57.40) and AA-LCR (64.6). Qwen3.6-35B-A3B is ahead on coding, science and knowledge rows, with 73.4 on SWE-bench Verified shown as a reported score rather than a re-run, while Dory leads it on IFBench, AA-LCR and the τ³-bench Banking test. The AA-style rows are NVIDIA's internal estimates and may differ from official Artificial Analysis scores, and HokAI found no independent evaluation as of 20 October 2026.

The weights are the small part of the footprint. The BF16 file is 7.95 GB, the MXFP8 export 4.75 GB, the NVFP4 export 3.20 GB and the Q5_K_M GGUF 3.67 GB. Every benchmark above is for BF16: NVIDIA publishes no accuracy figures for the quantized exports. KV-cache memory is the other cost. The per-loop mode used in the card's launch commands stores a separate cache for each loop, so memory grows with the loop count, while last-loop caching keeps one cache per recurrent attention layer and holds the footprint at one-loop size. NVIDIA reports that accuracy stays close to per-loop mode with nearly unchanged output lengths, without publishing the figures.

Running it takes extra steps today. As of 20 October 2026 support sits in two open pull requests, vLLM 60149 and llama.cpp 30306, so the card tells vLLM users to install from the pull-request ref and llama.cpp users to build a pinned commit from a fork. NVIDIA's test hardware is H100, B200, B300 and DGX Spark, the card names Linux as the preferred operating system, and its llama.cpp walkthrough targets Apple Silicon Macs. Recommended sampling is temperature 1.0 and top-p 0.95. The loop count is chosen when the server starts, through n_recurrent_loops in vLLM or dory.num_loops in llama.cpp. If you would rather use models that run in today's local tools, start with How to Run AI Locally in 2026.

Input and output are English text only, and the card says nothing about images, audio or other languages. Reasoning is on by default: the bundled chat template turns thinking on unless told otherwise, carries a medium-effort switch the card never mentions, and by default strips reasoning text from turns before the latest user message. Tool calls use an XML format that vLLM parses with its qwen3_coder parser. The 262,144-token position limit in the config matches the 256K length NVIDIA says it validated, and training covered 43 programming languages alongside English.

The licence is the least restrictive part. OpenMDW-1.1 grants free permission to deal in the model materials without restriction and puts no obligation on outputs. Redistributors keep the licence text and origin notices, and your rights end if you sue someone claiming the model infringes a patent or copyright, unless you are answering a suit first brought against you. Hugging Face tags it license:other, so Hub licence filters will not group it with Apache or MIT. The card links safety, bias, explainability and privacy subcards, but none of those files exist in the repositories, and no red-team partner is named. Its own bias note says the document datasets skew toward male references and toward White as the most frequent ethnicity term, at 43 to 44 percent of ethnicity mentions.

Dory 1.0 4B fits developers who self-host a small English coding or agent model and can live with pre-release runtimes, and researchers who want to study test-time compute in latent space using one set of weights. It is the wrong pick for teams that need image input (try Ministral 3 3B or NVIDIA's Nemotron 3.5 Super VL 120B-A12B), multilingual output, or a hosted API (NVIDIA's Nemotron 3.5 Lightning lists hosted options). Teams with room for a larger model can look at Qwen3.6-27B or Gemma 4 12B. Browse the wider field in the open-source model list, the NVIDIA model directory and the older Nemotron 3 Nano 30B A3B. Dory sits in the 128k to 400k context bucket; for what else runs locally today, see Best AI Models You Can Run Locally for Coding in 2026, or work through Which LLM Should You Use? first.

Pricing

The weights are free under the permissive licence, so the bill is your own hardware. HokAI found no hosted per-token endpoint for Dory when this page was written. For a per-token option from the same publisher, see Nemotron 3.5 Lightning.

Key Features

• Adjustable loop count: One checkpoint can run its shared recurrent block 1, 2 or 4 times per token (llama.cpp also accepts 3), chosen at server launch with no retraining or model swap.
• Last-loop KV caching: An optional mode keeps a single cache per recurrent attention layer, so KV memory stays at one-loop size however many loops run.
• Native NVFP4 pre-training: Built on a normalized-GPT backbone, the model was trained in NVIDIA's 4-bit NVFP4 format for the first ~4.5T tokens before finishing in BF16.
• Hybrid attention layout: Sliding-window attention (1,024 tokens) covers three of every four attention layers, the fourth is full attention, and grouped-query attention uses 4 key-value heads against 16 query heads.
• Four release formats: BF16, MXFP8, NVFP4 and Q5_K_M GGUF; on DGX Spark at batch size 1, NVIDIA measured NVFP4 at 2.17x to 2.44x and MXFP8 at 1.69x to 1.84x the speed of BF16.
• Reasoning and tool calling: Thinking is on by default and tool calls use an XML format parsed by vLLM's qwen3_coder parser; reinforcement learning covered multi-step tool use and structured output.

Pros

• Wins on the two jobs small local models are most often given, agentic coding and strict instruction following, against the same-size Qwen3.5-4B in NVIDIA's table (vendor-reported).
• The licence is short and permissive, so legal review is quick and outputs carry no obligations.
• One deployment can serve cheap and hard requests by changing a launch flag, instead of keeping two model sizes loaded.

Cons

• Needs pre-release runtime builds: vLLM and llama.cpp support lives in open pull requests, so teams must pin a branch or commit.
• English text only, with no image, audio or multilingual support.
• Quantized exports have no published accuracy figures, and the repositories carry no safety, bias or privacy subcards.
• More loops are not always better: graduate-level science and Humanity's Last Exam slip at the highest loop count in NVIDIA's own table.

Benchmarks

• IFBench: 77.9% vendor-reported · 20 Oct 2026 — How precisely the model follows detailed instructions, % passing.
• GPQA Diamond: 72.4% vendor-reported · 20 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
• SWE-bench Verified: 59.5% vendor-reported · 20 Oct 2026 — Real GitHub issues fixed end to end, % solved.
• Terminal-Bench 2.1: 18.1% vendor-reported · 20 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
• Humanity's Last Exam: 10.9% vendor-reported · 20 Oct 2026 — Expert-written questions across many fields, % correct.

A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.

Visit Nemotron Labs Dory 1.0 4B Official Page