NVIDIA-Nemotron-3.5-Lightning-30B-A3B

NVIDIA-Nemotron-3.5-Lightning-30B-A3B is an open language model from NVIDIA, released in two weight variants: BF16 (reference weights for customization and research) and optimized NVFP4 (for fast inference and deployment). The model is positioned as the “fastest 30B MoE” for specialized tasks within always-on agents: with 30 billion total parameters, only 3 billion are activated per token, which ensures high inference throughput. Native context length is 262,144 tokens, with the ability to extend to 1 million tokens.

The model is based on a hybrid Mixture-of-Experts architecture in which Mamba-2 and Latent MoE layers alternate, while classic Attention layers are used only at a ratio of approximately 1:3. This hybrid attention solves the key problem of long context: the self-attention mechanism scales quadratically in memory and computation with sequence length, whereas Mamba-2 (a state space model) scales linearly. Combining both approaches allows the model to process context without explosive cost growth while maintaining accuracy on salient details in requests. In this architecture, Latent MoE means that expert routing and expert computation occur in a hidden (latent) representation space rather than in the full dimensionality of the model layers. This reduces routing and computation overhead and enables faster inference computation.

Compared with the previous model in the Nemotron-3-Nano-30B-A3B series, the architecture remains the same; however, Lightning 3.5 demonstrates substantially higher results on agentic and coding benchmarks. For example, on SWE-bench Verified, Lightning scores 51.56 versus 34.08 for Nano; on SWE-bench Multilingual, 39.33 versus 14.07; on Terminal-Bench 2.1, 24.58 versus 8.29; and on AA-LCR (long context), 52.00 versus 32.75. This is achieved through higher-quality post-training (SFT and RL on synthetic data focused on tool calling, structured outputs, and multi-step scenarios).

Thanks to its hybrid long-context architecture and low activation cost, the NVFP4 variant is recommended for deployment in AI agent systems, chatbots, RAG systems, and applications where strict instruction following is critical. The model is well suited for always-on agents performing specialized tasks, as well as for local inference.

Announce Date: 01.08.2026
Parameters: 32B
Experts: 128
Activated at inference: 3B
Context: 263K
Layers: 52, using full attention: 6, using no attention: 23
Attention Type: Hybrid Attention
Mamba Type: Mamba 2
Developer: NVIDIA
Transformers Version: 4.57.6
vLLM Version: >=0.28.0
License: openmdw-1.1

Use our pre-built public endpoints for free to test inference and explore NVIDIA-Nemotron-3.5-Lightning-30B-A3B capabilities. You can obtain an API access token on the token management page after registration and verification.

An endpoint is a server with pre-installed software that automatically downloads neural network weights and launches an OpenAI-compatible API to handle incoming requests.

A private endpoint is created within the user's project, belongs exclusively to the user, and is billed per server configuration in accordance with the infrastructure pricing.

We recommend creating private endpoints when you need to:

• achieve maximum endpoint performance,
• access the full context window for processing large requests and conversations,
• ensure maximum security by processing data in a dedicated, isolated environment.

Contact our dedicated neural networks support team at support@immers.cloud or send your request to the sales department at sale@immers.cloud.