Model library / nvidia
NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr is a 673M mixture-of-experts open-weight model published by nvidia on Hugging Face. It uses the LiLiCorrDraftModel architecture with 5 layers. The published configuration allows a context of 262,144 tokens. The weights are released under the other licence.
Mixture of experts, count not published
Context length in tokens, as published
Weights at the released precision
Licence declared on the repository
Overview
NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr is a 673M mixture-of-experts open-weight model published by nvidia on Hugging Face. It uses the LiLiCorrDraftModel architecture with 5 layers. The published configuration allows a context of 262,144 tokens. The weights are released under the other licence.
At the released precision the weights occupy about 1 GiB, before any key-value cache. That fits on a single RTX 4090 24GB. The tables below work the memory out for common cards.
Specification
| Repository | nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr |
| Publisher | nvidia |
| Published | 12 October 2026 |
| Architecture | LiLiCorrDraftModel |
| Model type | qwen3 |
| Parameters | 673,450,538 |
| Layers | 5 |
| Hidden size | 4,096 |
| Attention heads | 32, 2 key/value heads |
| Context length | 262,144 tokens |
| Vocabulary | 131,072 tokens |
| Quantisation | modelopt_fp4 |
| Licence | other |
| Base model | nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB |
Memory
Parameter count multiplied by the bytes each format uses. The key-value cache comes on top.
| Precision | Weights | Notes |
|---|---|---|
| bf16 | 1.3 GiB | Full precision as released |
| fp8 | 0.6 GiB | 8-bit, near-lossless on most models |
| int4 | 0.3 GiB | 4-bit, smallest footprint |
The cache costs 5 KiB per token at 16-bit, so the full 262,144-token context of one request needs about 1.3 GiB. Concurrency multiplies that number, not the weights.
That figure is an upper bound: this model uses sliding-window attention with a 1,024-token window on some layers, so long requests cache less than this.
Hardware
At the released precision, with 90% of the card given to the server and 2.5 GiB kept for the runtime. The context column is what is left for the cache on that setup.
| GPU | Cards needed | Free for cache | Context it holds |
|---|---|---|---|
| RTX 4090 24GB | 1 | 18.3 GiB | the full 262,144 tokens |
| RTX 5090 32GB | 1 | 25.5 GiB | the full 262,144 tokens |
| L40S 48GB | 1 | 39.9 GiB | the full 262,144 tokens |
| A100 80GB | 1 | 68.7 GiB | the full 262,144 tokens |
| H100 80GB | 1 | 68.7 GiB | the full 262,144 tokens |
| RTX PRO 6000 96GB | 1 | 83.1 GiB | the full 262,144 tokens |
| H200 141GB | 1 | 123.6 GiB | the full 262,144 tokens |
Serving
A vLLM launch line for the shape above. Check the model card for a runtime the publisher recommends.
vllm serve nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr \ --max-model-len 262144 \ --gpu-memory-utilization 0.90
Questions
The weights take about 1 GiB at the released precision. That fits on one RTX 4090 24GB, which leaves roughly 18 GiB for the key-value cache.
The repository declares other. Read the licence text before commercial use: the name alone does not tell you what is allowed.
The configuration allows 262,144 tokens. Whether the whole window is usable depends on the memory left for the cache, which the hardware table works out per card.
This page is a reference, and Aiqre does not serve this model today. We deploy open-weight models on dedicated EU hardware on request, so if you want an OpenAI-compatible endpoint for it, email [email protected] with the model name and roughly what volume you expect.
Also in the library
A 0.9B decision model that returns probabilities for multiple questions in a single forward pass, not a text generator.
Limite 1B, Violetto is a 1-billion-parameter dense autoregressive transformer aimed at high-throughput mathematical reasoning.
K2-Horizon-0.9B is a 1.1 billion-parameter dense decoder-only model with a 128 k token context window.
Nanbeige4.2-3B is a compact agentic language model designed for reasoning, tool use and personal-assistant tasks.
Hosting
Tell us the model and roughly what volume you expect. We reply with a price and an OpenAI-compatible endpoint, with your prompts never stored.