Nemotron 3.5 Lightning

Nemotron 3.5 Lightning is a hybrid mixture-of-experts with 30B total and 3B active parameters. Its published config shows 52 layers: 23 Mamba-2 state-space layers, 23 MoE feed-forward layers and 6 attention layers, plus one multi-token-prediction layer. Each MoE layer holds 128 routed experts and one shared expert, with 6 routed experts active per token; the few attention layers use grouped-query attention (32 query heads, 2 KV heads). Because most of the stack is state-space rather than attention, the attention-type field is left null, as it is for the rest of the Nemotron 3 family. Vocabulary is 131,072 tokens. Pre-training ran on more than 20 trillion tokens with an NVFP4 recipe, followed by continued pre-training for the MTP layer, supervised fine-tuning (code, math, science, tool calling, structured output, long-range retrieval) and multi-environment GRPO reinforcement learning. Part of the post-training data was distilled from open teacher models including GPT-OSS-120B and Qwen3-235B. The config ships with a 262,144-token position limit; NVIDIA's 1M-token recipes raise it explicitly and call for 8x H100, a GB200 or a B200. Training compute is not disclosed.

Coding (6.8) rests on NVIDIA-measured SWE-bench Verified 51.56, SWE-bench Multilingual 39.33 and Terminal-Bench 2.1 24.58: useful for routine repository tasks and code review, but clearly behind Qwen3.6-35B-A3B (70.12 / 63.40 / 44.38) and Gemma 4 26B A4B (57.40 on SWE-bench Verified). Reasoning (7.5) is anchored by MMLU-Pro 81.94 and GPQA Diamond 75.44, solid for 3B active parameters but eight points below Qwen on GPQA; HLE 11.72 confirms it is not a hard-reasoning model. Math (7.0) has no dedicated published score (no AIME row), so it is scored conservatively. Agentic work (7.0) is the design target: PinchBench 85.37 is close to Qwen's 88.07, and BrowseComp 36.97 and GDPval-AA-V2 832 beat Gemma 4, but τ³-bench Banking 9.28 is weak, as it is for the whole class. Instruction following (7.5) is a relative strength, with IFBench 71.88 ahead of Qwen's 63.71. Function calling (7.5) reflects harness-specific tool-call training and a documented tool parser. Long context (7.0) offers a real 1M window, but AA-LCR 52.00 trails Qwen (61.06) and Nemotron 3 Super (58.44). Multilingual (6.0) covers English plus Spanish, French, German, Italian and Japanese. Creative writing (5.5) is not a focus. Safety calibration (6.5) is left largely to the deployer. The model is text-only, so vision and OCR score 0, and it has no built-in live data (0).

Every figure is vendor-reported: NVIDIA measured all the models in this table, competitors included, under its own NeMo Gym / NeMo Evaluator harness, and notes the numbers may differ from each vendor's self-reported scores. NVIDIA publishes the recipes to reproduce them. Scores are for the BF16 checkpoint; the NVFP4 checkpoint lands within about three points on every percentage row (SWE-bench Verified 52.80, MMLU-Pro 81.62, AA-LCR 49.19). MMLU-Pro, GPQA Diamond, SWE-bench Verified, Terminal-Bench 2.1 and HLE map to the catalog's compare columns; IFBench (not IFEval), AA-LCR (not MRCR) and τ³-bench Banking (not τ-bench) stay in this table only, to avoid mismatched compare rows. No independent leaderboard score is cited yet.

No independent throughput or first-token measurement is available yet, so the speed fields are null. NVIDIA's own claims, which are vendor-reported: up to 4x the output speed of similar-sized models, and 10,000 PinchBench tasks completed 30% faster than Qwen3.6 35B at comparable accuracy. The mechanics behind those claims are concrete: only 3B parameters are active per token, the Mamba-2 layers keep long-context cost down, and three speculative-decoding options ship with the model (built-in MTP for medium-to-high concurrency, the DSpark drafter for DGX Spark and low-concurrency data-centre serving, and DFlash). Reasoning is on by default; turning it off with enable_thinking=False is the main latency lever for short tool-call turns.

NVIDIA's business is GPUs, not tokens, so every price field is null. Self-hosting costs only hardware: the NVFP4 build runs on a single RTX 5090 or DGX Spark, and BF16 fits one 80GB H100 at a 256K context. Metered access runs through third-party hosts at their own prices.

Open weights on Hugging Face in three NVIDIA checkpoints: BF16 (reference weights for customization), NVFP4 (NVIDIA's recommended build for production inference) and Base (pre-trained, for fine-tuning), plus the DSpark and DFlash drafter models and an official GGUF from ggml-org. NVIDIA's validated recipes run BF16 on one H100 or A100 80GB at a 256K context, and the full 1M context on 8x H100 (tensor plus expert parallel), a GB200 or a B200; NVFP4 runs on DGX Spark, H100 and GB200, with Ampere covered through W4A16, and TensorRT-LLM recipes sit on the NVFP4 card. For local use, Ollama (ollama run nemotron-3.5-lightning), llama.cpp and LM Studio all have recipes; NVIDIA names the GeForce RTX 5090 as its smallest discrete-GPU target, and Ollama sizes context to available memory (4K under 24GB, 32K at 24 to 48GB, 256K above 48GB). NVIDIA also lists Jetson, RTX PCs and DGX Station as targets. Managed routes: the NIM container nvcr.io/nim/nvidia/nemotron-3.5-lightning-30b-a3b, Amazon SageMaker JumpStart, Google Cloud Gemini Enterprise Agent Platform, Microsoft Foundry and OCI Enterprise AI, plus hosted inference from Baseten, Bitdeer AI, BlackBox AI, CoreWeave, Crusoe, DeepInfra, Fireworks AI, FriendliAI, GMI Cloud, Modal, Nebius and Together AI, OpenRouter, and a ModelScope mirror.

NVIDIA frames the release under its Trustworthy AI policy and ships Model Card++ subcards for bias, explainability, privacy and safety rather than a named safety framework, so that field is null. The card documents a training-data bias analysis and filters that removed teacher-model reasoning traces aligned with political entities or nationalistic narratives, and it asks deployers not to remove the model's safety guardrails without an equivalent replacement. Self-hosting keeps prompts on your own infrastructure, so NVIDIA never sees inputs; the free build.nvidia.com endpoint is NVIDIA-hosted and governed by NVIDIA's terms of use. No compliance certifications attach to an open-weights release. Moderation is the deployer's job: NVIDIA's Nemotron Safety models and NeMo Guardrails are the matching tools, and community fine-tunes with the safety tuning removed already circulate on Hugging Face. OpenMDW-1.1 permits commercial use but is new enough to warrant a legal read.

Serving support landed on day one across vLLM, SGLang and TensorRT-LLM, with NVIDIA NIM for containerized deployment and Ollama, llama.cpp, LM Studio, Unsloth and Exo for local use. NVIDIA lists agent-harness partners including Cline, Kilo Code, OpenHands, OpenCode, Factory AI, Hermes Agent, OpenClaw and LangChain, and post-training partners including Prime Intellect, Thinking Machines Lab and distil labs. Named enterprise customizers are CrowdStrike (cybersecurity), Harvey with Trajectory (legal), CodeRabbit with Baseten (code review), Lila Sciences and Fastino Labs. Adoption is strong for a six-week-old model: on Hugging Face the NVFP4 checkpoint has about 1.59 million all-time downloads and BF16 about 0.70 million, with dozens of community GGUF, MLX and FP8 builds, including a Red Hat FP8 version. NeMo Switchyard, released alongside it, makes Lightning a routing target next to frontier models, and it sits in the Nemotron 3 family beneath Nemotron 3 Super and Nemotron 3 Ultra.