NVIDIA NIM on DGX Spark: which NIMs support GB10, what the license allows, and NIM vs your own vLLM

As of 7 November 2026, NVIDIA's NIM documentation lists about 20 microservices that run on the DGX Spark's GB10 chip: nine model-specific LLM NIMs plus the model-free vLLM container, three older Spark-only LLM containers, three embedding NIMs, four speech NIMs and one two-node vision NIM for GLM-5.3-Flash. Developer Program membership covers NIM for development and testing on up to 16 GPUs at no cost; serving real users needs an NVIDIA AI Enterprise license. I found no published tok/s for a NIM on a Spark, and NVIDIA's own release notes record a 20% throughput gap against open-source vLLM for one Spark NIM.

Which NIM microservices support DGX Spark?

Mostly NVIDIA's own Nemotron models, plus a few Llama, Qwen and gpt-oss builds. Support is spread over four separate support matrices; this table combines them.

Read 7 November 2026. "Verified" means the matrix lists NVIDIA-GB10 among verified GPUs without a GB10-specific tuned profile. The LLM matrix references NIM 2.1.4 and model-free container 2.1.5 and shows no page date. NVIDIA's NGC DGX Spark collection (updated 16 March 2026) lists five NIMs: Multi-LLM NIM, Nemotron-3-Nano-30B-A3B, Nemotron-Nano-9B-v2, Llama-3.1-8B and Qwen3-32B.

The list has been growing. In January 2026 a forum user complained that NIM pages for Qwen3-Next, Llama and Mistral had "no DGX Spark identity", and in the same month owners reported the Parakeet speech NIM as x86-only. By November, the NIM 2.x entries in the matrix list GB10 next to datacenter GPUs, and the playbook's default model ships as one multi-platform container with no "-dgx-spark" build. For speech models outside NIM, see the speech-to-text guide; for embeddings, the embeddings and vector search guide.

Do you need a license to run NIM on a Spark?

Not for development; yes for production. The two routes come from NVIDIA's NIM FAQ and the AI Enterprise DGX Spark quick start (updated 4 November 2026).

The FAQ defines production as business transactions or any non-testing activity, such as serving real end users. The quick start says NIM containers can be reached through either Developer Program membership or AI Enterprise, and that the DGX Spark entitlement applies only if you purchased it, requested an evaluation or received an entitlement certificate. eurokommerz found no perpetual license for the Spark edition. This is a summary of NVIDIA's pages, not legal advice.

The practical reading: an engineer evaluating a model on one Spark is inside the free tier. A Spark that answers colleagues' questions every day is production, and needs AI Enterprise. Open-source vLLM, SGLang and llama.cpp carry no such split, which matters more than any speed difference.

How do you start a NIM on a Spark?

Get an NGC API key, log Docker in to nvcr.io, and run the container with a cache volume and port 8000. The commands below are NVIDIA's from the nim-llm playbook, with the variables filled in for its default model; the NIM_KVCACHE_PERCENT line is the GB10 setting from the LLM support matrix, and the port is bound to localhost.

export NGC_API_KEY="<YOUR_NGC_API_KEY>"
echo "$NGC_API_KEY" | docker login nvcr.io --username '$oauthtoken' --password-stdin

export IMG_NAME=nvcr.io/nim/nvidia/nemotron-3.5-lightning-30b-a3b:latest
export MODEL_ID=nvidia/nemotron-3.5-lightning-30b-a3b
export CONTAINER_NAME=nemotron-lightning
export LOCAL_NIM_CACHE=~/.cache/nim
export LOCAL_NIM_WORKSPACE=~/nim-workspace
mkdir -p "$LOCAL_NIM_CACHE" "$LOCAL_NIM_WORKSPACE"

docker run -it --rm --name=$CONTAINER_NAME \
  --gpus all \
  --shm-size=16GB \
  -e NGC_API_KEY=$NGC_API_KEY \
  -e NIM_MODEL_NAME=$MODEL_ID \
  -e NIM_SERVED_MODEL_NAME=$MODEL_ID \
  -e NIM_KVCACHE_PERCENT=0.75 \
  -e NIM_PASSTHROUGH_ARGS="--reasoning-parser nemotron_v3" \
  -v "$LOCAL_NIM_CACHE:/opt/nim/.cache" \
  -v "$LOCAL_NIM_WORKSPACE:/opt/nim/workspace" \
  -p 127.0.0.1:8000:8000 \
  $IMG_NAME

curl -s http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "nvidia/nemotron-3.5-lightning-30b-a3b",
       "messages": [{"role": "user", "content": "Summarize GPU computing in one sentence."}],
       "max_tokens": 128, "chat_template_kwargs": {"enable_thinking": false}}'

The playbook puts the job at about 30 minutes, longer on the first run while the model downloads, and needs 19 to 63 GB of disk depending on precision. It works on both the 64 GB and 128 GB Spark configurations. The API is OpenAI-compatible, so anything that talks to vLLM talks to this. Reach it from your laptop through an SSH tunnel rather than publishing port 8000.

How fast is NIM on a Spark compared with vLLM?

Nobody has published a NIM tok/s figure on a Spark that I could find, and NVIDIA's only Spark comparison favors open-source vLLM. The NIM 1.14.0 release notes list "a 20% degradation in throughput observed for vLLM at FP8 precision compared to OSS vLLM at concurrency=10" for the Llama 3.1 8B DGX Spark NIM, at input/output lengths of 2,000/200, 200/2,000 and 1,000/1,000.

Current GB10 profiles are vLLM-based (the Nemotron 3.5 Super VL profile is named vllm-nvidia-gb10-nvfp4-tp1-pp1-throughput), so my expectation, not a measurement, is that a NIM lands near stock vLLM on the same checkpoint and settings. For the playbook's model, Kubesimplify reports Nemotron 3.5 Lightning on the stock vllm/vllm-openai:v0.27.1 image with NVFP4 weights and FP8 KV at about 108 tok/s single-stream on a short prompt and 88 to 90 on a long one, with a 3-token speculative draft. The same post measured Ollama v0.32.9 at 72 to 87 tok/s on a Q4_K_M build. vLLM took just under 2 minutes to initialize against 27.3 seconds for Ollama, and at --gpu-memory-utilization 0.80 held 21.5 GB of weights and 75 GB of KV cache.

A NIM picks its own profile and may not enable the same speculative decoding, so measure before you assume. GPUwerk's own vLLM figures, for other models, are on the benchmarks page.

NIM or your own vLLM or SGLang?

NIM if you need NVIDIA support and run one of its listed models; your own engine for everything else.

• Choose NIM for a Spark you own that serves a Nemotron, Llama or Qwen NIM to a team, under an AI Enterprise license, where a supported, version-pinned container is worth more than the last 20% of speed.
• Choose vLLM or SGLang for new models (the reported DeepSeek V4 and GLM-5.3 runs on Sparks use community builds, see the multi-Spark guide), for speculative decoding you tune yourself, for commercial serving without a per-GPU subscription, and on any rented instance without Docker. The engine comparison guide gives Spark tok/s for the three engines.
• Choose llama.cpp or Ollama for GGUF builds, the only way some models fit one Spark.

My recommendation for most readers: run the NIM playbook once on Developer Program terms to see what NVIDIA tuned, then serve with vLLM unless you are buying AI Enterprise anyway. The model-free NIM container narrows the coverage gap, but the license question stays.

Which build.nvidia.com/spark playbooks matter here?

The DGX Spark playbook index lists about 30 playbooks, with separate groups for those that run on both the 64 GB and 128 GB Spark and those for 128 GB only. For serving models, the ones worth reading are nim-llm (this NIM), vllm, sglang and llama-cpp; for multiple nodes, connect-multiple-sparks, vllm-cluster and nccl, which the two-Spark cluster guide walks through; and nvfp4-quantization and lm-studio, both listed as 128 GB only. No speech or Riva playbook appears on the index; the speech NIMs are documented only in the NIM docs. NVIDIA's NIM developer page also offers hosted NIM API endpoints for prototyping, which let you try a model before downloading anything. NVIDIA’s project tool for an owned Spark is covered in AI Workbench on a DGX Spark.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• Run an LLM with NVIDIA NIM (nim-llm playbook), NVIDIA build.nvidia.com, 2026-09-30
• nim-llm playbook: instructions, NVIDIA build.nvidia.com
• DGX Spark playbooks, NVIDIA build.nvidia.com
• Support Matrix for NIMs (NVIDIA NIM for LLM and VLM), NVIDIA Docs
• Release Notes for NVIDIA NIM for LLMs 1.14.0, NVIDIA Docs, 2026-07-23
• Support Matrix, NVIDIA NIM for Vision Language Models, NVIDIA Docs, 2026-09-30
• Support Matrix for NVIDIA NeMo Retriever Embedding NIM, NVIDIA Docs, 2026-11-05
• NVIDIA ASR NIM Support Matrix (26.10.0), NVIDIA Docs, 2026-11-06
• NVIDIA NIM FAQ, NVIDIA Developer Forums
• NVIDIA AI Enterprise: DGX Spark Quick Start Guide, NVIDIA Docs, 2026-11-04
• NVIDIA DGX Spark collection, NVIDIA NGC Catalog, 2026-03-16
• NVIDIA AI Enterprise DGX Spark Licence Explained, eurokommerz, 2026-11-05
• Running Nemotron 3.5 Lightning on DGX Spark, Kubesimplify, 2026-08-11
• DGX Spark + Qwen3-Next-80B: Proven Performance, But Missing Clear Path to NIM, TensorRT-LLM & Web UIs, NVIDIA Developer Forums, 2026-01-16
• NIM for Developers, NVIDIA Developer