Skip to content
Nemotron 3.5 Lightning logo

Nemotron 3.5 Lightning

NVIDIA

Key facts · verified Oct 29, 2026

Nemotron 3.5 Lightning is a NVIDIA model released on 11 Aug 2026. Pricing: Free — open weights, self-hosted (no first-party API pricing). Its context window is 1M tokens (256k in NVIDIA’s single-H100 config). It ranks #217 on LMArena's text leaderboard (Elo 1353, read 1 Nov 2026). Source: huggingface.co.

NVIDIA’s Nemotron 3.5 Lightning, released 11 August 2026: a 30-billion-parameter hybrid Mamba-2 / MoE / attention model with just 3 billion parameters active per token, published under the commercially-approved OpenMDW License 1.1. Inference cost tracks a 3B model while capacity tracks a 30B one, which is what lets it run agentic workloads on a single H100 or A100 80GB, including in a workstation. Context reaches 1M tokens, though NVIDIA’s own single-GPU configuration uses 256k.

Pricing & specs checkedSource: vendor documentation · checked Oct 29, 2026LMArena: 1353 Elo (rank 217) · source · retrieved Nov 1, 2026
Pricing

Free — open weights, self-hosted (no first-party API pricing)

Context

1M tokens (256k in NVIDIA’s single-H100 config)

LMArena Elo

1353

LMArena rank

#217

Key Features

Open Weights Single-GPU Mamba-MoE Hybrid 3B Active Params Tool Calling Toggleable Reasoning

Full Specifications

Free — open weights, self-hosted (no first-party API pricing)
Identity
DeveloperNVIDIA
ReleasedAug 2026
StatusGA
LicenceOpenMDW License Agreement 1.1 (approved for commercial use)
Self-hostableYes — 1x H100 80GB or 1x A100 80GB; also Blackwell (GB200, B200) and Hopper (H200)
Cost
Input priceSelf-hosted — no first-party API price
Output priceSelf-hosted — no first-party API price
Cached inputNot applicable
Batch discountNot applicable
Free tierYes — weights are free to download and run
Capacity
Context window1M tokens (256k in NVIDIA’s single-H100 config)
Max outputNot documented
Long-context surchargeNot applicable
Capability
Vision inNo — text only
Audio inNo
Function callingYes — tool calling via vLLM
Structured outputYes
Extended reasoningYes — configurable on/off via chat template (enable_thinking)
Web searchNo
Code executionNo
Access
APINot offered first-party
Chat appNo
Cloud marketplacesNVIDIA NIM / build.nvidia.com
Fine-tuningYes — open weights permit it
Measured quality
LMArena Elo1353 (checked Nov 2026)
LMArena rankRank 217