Primary use stated or inferred cautiously from official documentation.
N
NVIDIA / nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16
Nemotron Labs Dory 1.0 4B
NVIDIA text model released on October 16, 2026, with 4 billion parameters and a recurrent architecture. It reuses a block of layers to adjust compute per token without switching checkpoints. English and a validated 256K-token context are documented. It does not generate images, audio or video.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 |
|---|---|
| Model type | Text and reasoning · Code |
| Access model | Open weights |
| License | OpenMDW 1.1 |
| Deployment | Self-hosted; Dory support pending stable releases |
| Release | 16 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text |
| Exit | Text |
| Context window | 262,144 tokens |
| maximum output | Not published |
| Reasoning | 1, 2 or 4 loops; 2 recommended |
| Published tools | Features |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · vLLM · llama.cpp |
Reasoning, coding and tool-using agents on your own infrastructure. NVIDIA recommends two loops as a starting point; one reduces compute per token and four may improve some tasks. Measure end-to-end latency and cost: more loops do not guarantee higher accuracy on every benchmark.
Dory support in vLLM and llama.cpp requires the revisions specified by NVIDIA, not an arbitrary stable release. The local dory alias is not a commercial API. The GGUF example’s 4,096-token context and 2,048 output tokens are not universal limits. We do not transfer BF16 results to quantized formats.
03 / CAPABILITIES
What it can do
Orientation
Reasoning, coding and tool-using agents on your own infrastructure. NVIDIA recommends two loops as a starting point; one reduces compute per token and four may improve some tasks. Measure end-to-end latency and cost: more loops do not guarantee higher accuracy on every benchmark.
Context
262,144 tokens · Not published
Tools and integration
Features
Access
Hugging Face · downloadable checkpoint
Input modalities
Text
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
Weights are available in BF16, MXFP8, NVFP4 and GGUF Q5_K_M under OpenMDW 1.1, with commercial and non-commercial use subject to its terms. We have not verified a hosted API tariff: numeric cost fields remain empty. Downloads do not remove hardware, memory, operating and security costs.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Ten NVIDIA-published metrics for BF16, not reproduced by Inferama. The primary figure uses two loops; the linked conditions also preserve one- and four-loop results. Terminal-Bench 2.1 uses Terminus-2. AA-LCR keeps the Qwen3-235B-A22B judge and GDPval-AA v2 uses specific anchors: these are not direct comparisons with other protocols. We do not present chart estimates of the AA index as official scores.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| SciCode | 34,94 | BF16 score · 2 loops | View source ↗ |
| SWE-bench Verified | 59,5 | BF16 score · 2 loops | View source ↗ |
| Terminal-Bench 2.1 (Terminus 2) | 18,1 | BF16 score · 2 loops | View source ↗ |
| Tau3-bench (Banking) | 12,89 | BF16 score · 2 loops | View source ↗ |
| GDPval-AA v2 (NVIDIA anchors) | 946 | BF16 Elo · 2 loops | View source ↗ |
| GPQA Diamond (sin herramientas) | 72,41 | BF16 score · 2 loops | View source ↗ |
| HLE (texto, sin herramientas) | 10,89 | BF16 score · 2 loops | View source ↗ |
| IFBench | 77,94 | BF16 score · 2 loops | View source ↗ |
| AA-LCR (NVIDIA, Qwen3 judge) | 71 | BF16 score · 2 loops | View source ↗ |
| AA-Omniscience (accuracy) | 10,08 | BF16 score · 2 loops | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
BF16 / MXFP8 / NVFP4 / GGUF Q5_K_M
Four formats of one model, not four independent releases. The reference identifier is BF16; official repositories for each format are linked in the sources. per_loop mode increases KV cache with each loop; last_loop keeps a single cache per recurrent layer.
07 / SOURCES