Strengths
- It fits in 24 GB of memory, so it runs on high-end GeForce RTX cards with enough memory (NVIDIA tested it on an RTX 5090), on DGX Spark and on RTX Spark laptops.
- Common local model tools support it: llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
- In NVIDIA's table it beats Gemma 4 26B A4B and Qwen3.6-35B-A3B on most agentic tests, for example 66.75 on BrowseComp against 26.30 and 48.74.
- Besides text it understands images and video, has a 256K token context, and reasoning can be switched on or off.
- The permissive OpenMDW 1.1 license allows commercial use and changes.
Weaknesses
- All results come from NVIDIA. On the AA-Omniscience knowledge test its accuracy is 16.90 against 22.17 for Gemma 4 26B A4B. It knows fewer facts from memory, so pair it with search for factual questions.
- It needs at least 24 GB of graphics card or unified memory, such as an RTX 5090 or DGX Spark. An ordinary office laptop will not do.
- Its main languages are English and code, then Spanish, French, German, Italian and Japanese. Slovak and Czech are not on the list.
- It is not on OpenRouter yet. The model is mainly meant for your own machine.
- NVIDIA sized the 24 GB budget for a 128K token context, so on a 24 GB card you will probably not get the full 256K.
Best for
- an AI agent on your own powerful PC
- sensitive data that must stay in the company
- web search and tool use without the cloud
- terminal tasks
Not ideal for
- ordinary office laptops
- texts in Slovak or Czech without testing first
How to use it in Europe
The weights are free to download from Hugging Face anywhere in Europe and run in Ollama, LM Studio or llama.cpp. The OpenMDW 1.1 license has no geographic restrictions. You need a PC with an NVIDIA graphics card and at least 24 GB of memory, such as a GeForce RTX 5090 or DGX Spark. The model runs entirely on your hardware, so you send no data to an AI provider. If the agent searches the web, only its search queries leave the machine.
Benchmark results
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 86.43% | huggingface.co ↗ (vendor-reported), October 2026 |
| Terminal-Bench 2.1 | 50.98% | huggingface.co ↗ (vendor-reported), October 2026 |
| BrowseComp | 66.75% | huggingface.co ↗ (vendor-reported), October 2026 |
| PinchBench | 87.82% | huggingface.co ↗ (vendor-reported), October 2026 |
| MCP-Atlas | 52.20% | huggingface.co ↗ (vendor-reported), October 2026 |
Benchmarks cover only a narrow slice of ability and vendors often report them at their best settings. Treat them as a rough guide.
More models from this vendor
- Nemotron 3.5 Super VL: NVIDIA's open model for business AI agents that understands text, images and video and is very fast and cheap to run.
Similar models from other vendors
- Gemma 4 31B (Google): Google's largest open model, which you can run on your own computer or server when data must not leave the company.
- Gemma 4 26B A4B (Google): Google's faster open model for your own server. Close to the larger Gemma 4 31B in quality, but quicker to respond.
- Muse Glimmer 30B (Meta): Meta's small open model under Apache 2.0 that runs on a single graphics card, even offline.
- Llama 4 Scout (Meta): The smaller Llama 4 from April 2025, very cheap through providers. Its license grants no rights to people or companies in the EU.
- Ministral 3 14B (Mistral AI): Mistral's small open model under Apache 2.0 that runs on a single GPU, even without an internet connection.
- GPT-6 Luna (OpenAI): OpenAI's cheapest model, made for high volumes of simple tasks, sorting text and low-cost chatbots.