NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4
Support summary
NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 is a checkpoint published by NVIDIA, built on the NemotronHForCausalLM architecture. Official recipes or cookbooks cover vLLM and SGLang, without pinning a stable release. TensorRT-LLM lists its architecture, but not this exact checkpoint.
- Documented quantization
- NVFP4
- Documented features
- Multi-token prediction, Reasoning parsers, Speculative decoding, Tensor parallelism, and Tool calling
- Configured context
- 262,144 tokens
- License
- openmdw-1.1
Hardware, quantization and features count only statements about this checkpoint; engine-wide capabilities are listed with each release.
Hugging Face model card
- Publisher
- NVIDIA
- Task
- text generation
- License
- openmdw-1.1
- Library
- transformers
- Architecture
- NemotronHForCausalLM
- Weight tensor elements
- 302.8B
- Tensor types
- BF16, F8_E4M3, F32, U8
- Quantization
- modelopt
- Configured context
- 262,144 tokens
- Languages
- en, fr, es, it, de, pt, ja, ko, hi, ar, zh, he
- Weight access
- Public
- Repository created
- 2026-06-03
Publisher deployment guidance
- vLLM0.22.0 container
The publisher NVFP4 example uses vllm/vllm-openai:v0.22.0 on 4 B200 GPUs. Context, parallelism and kernel settings are specific to that recipe.
Source: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 · publisher model cardReviewed - SGLang0.5.13 container
The publisher reports the NVFP4 container example tested on 4 B200 GPUs; this directory has not repeated that test.
Source: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 · publisher model cardReviewed - TensorRT-LLM1.3.0rc17 container
Publisher NVFP4 recipe for 4 B200 GPUs. This release candidate is outside the stable-release comparison.
Source: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 · publisher model cardReviewed
Model notes and sources
- nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 · publisher model card
- nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 · Hugging Face metadata
- nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 · publisher configuration
- vllm · models/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16.yaml
- sglang · docs/src/snippets/autoregressive/nemotron3-ultra-deployment.jsx
- sglang · docs/cookbook/autoregressive/NVIDIA/Nemotron3-Ultra.mdx
- Reviewed
Engine support
3 engines
| Engine | Release | Checkpoint support | Evidence scope | Last known supported |
|---|---|---|---|---|
| vLLM | Varies by configurationRecipe / cookbook | Documented recipeSource: vllm · models/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16.yamlReviewed | No supported release indexed | |
Features, dependencies and evidence for vLLM 0.28.0 | ||||
| SGLang | Version not pinnedRecipe / cookbook | Documented recipeSource 1: sglang · docs/cookbook/autoregressive/NVIDIA/Nemotron3-Ultra.mdxSource 2: sglang · docs/src/snippets/autoregressive/nemotron3-ultra-deployment.jsxReviewed | No supported release indexed | |
Features, dependencies and evidence for SGLang 0.5.19 | ||||
| TensorRT-LLM | 1.2.1Newest indexed | Architecture listed; exact checkpoint unknown | No supported release indexed | |
Features, dependencies and evidence for TensorRT-LLM 1.2.1 | ||||