AI Trend Notifier
EN한
← wiki

$ cat wiki/models/nemotron-3-5-lightning.md

Nemotron 3.5 Lightning

modelupdated 2026-08-12created 2026-08-12
nvidiaopen-weightsmoemambaagentslocal-inferencelong-contextdistillationrouting

Spec

AttributeValue
DeveloperNVIDIA
Released2026-08-11
Announced2026-08-11
Context window1M
Pricingunknown
LicenseOpenMDW-1.1
AvailabilityHugging Face (BF16 + NVFP4 open weights), NVIDIA NGC, build.nvidia.com, OpenRouter, DeepInfra, Together AI, Baseten, FriendliAI
Catalogue idnvidia/nemotron-3.5-lightning
30B total / 3B active hybrid Mixture-of-Experts, described as **interleaved
Mamba-2 and MoE layers with select Attention layers**
(source).
It is distilled from NVIDIA's frontier Nemotron 3 Ultra and developed with
the Nemotron Coalition, and ships with multi-token prediction and
DFlash speculative decoding
(source).

Data cutoffs as reported from the model card: pre-training September 2025, post-training May 2026 (source).

Pricing is unknown rather than free, on the same reading as Muse Glimmer: the weights are open and no first-party list price was published in anything read. Third-party serving prices exist and disagree with each other — DeepInfra at $0.05/M input · $0.20/M output at launch, some listings at $0.00/$0.00 free tiers, and OpenRouter carrying both a paid and a :free route (source). Those are resellers' prices for open weights, not NVIDIA's, and CLAUDE.md's rule for a catalogue price applies: any figure inside the provider range is real. The Catalogue id row is present so the daily spec-check Action can resolve one — it is copied from the OpenRouter listing URL, not guessed.

Release Date

2026-08-11, weights the same day on Hugging Face (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 and the -NVFP4 variant) and on NVIDIA NGC (source).

Shipped alongside NeMo Switchyard, an open-source model-routing library — the two were announced together and the pairing is the point rather than a coincidence of timing (see ## Compared To).

Benchmarks

NVIDIA's own figures except where the attribution says otherwise. No first-party benchmark table was read — the sandbox could not reach NVIDIA's blog or the model card (source).

MeasureFigureWhose number
Agent productivity (PinchBench)86.5%NVIDIA
Output speedup to 4x vs. comparable dense modelsNVIDIA
Agentic task completion timeup to 30% faster vs. classNVIDIA
Artificial Analysis Intelligence Index24Artificial Analysis (third party)
Two limits. **PinchBench appears in no sources/evals/ snapshot this repo
holds**, so there is nothing local to check 86.5% against — the same gap recorded
on Muse Glimmer for MCP Atlas and on Grok Imagine Image 2.0
for the image-edit board. And the **Artificial Analysis figure of 24 was read from
a search extract, not from a snapshot**: this repo's own Artificial Analysis
captures are produced by the weekly Action, and the most recent
(source) predates this
model by nine days, so it carries no row for it.

The claim shape is worth naming: three of the four figures are about speed or cost, not capability. That is consistent with a model positioned as the cheap step inside someone else's workflow rather than as a frontier competitor.

Use Cases

NVIDIA's stated framing is explicitly subordinate: a frontier reasoning model — Nemotron 3 Ultra, or GPT-5.6, which NVIDIA names — plans and orchestrates a workflow, while a smaller specialised model such as this one performs targeted steps: code review, tool use, security-alert monitoring, answering billing questions (source).

The 1M context window with 3B active parameters is the combination that makes that role affordable: long-horizon agent state can stay in context without paying dense-30B compute per token.

Compared To

ModelParamsActiveContextLicenceWeights
Nemotron 3.5 Lightning30B MoE3B1MOpenMDW-1.1released
Muse Glimmer30B dense30B131,072Apache 2.0released
Laguna S 2.1118B MoE8B1MOpenMDW-1.1released
Gemma 4 12B12B dense12Bunknownunknownreleased
The first two rows landed one day apart at the same parameter count, and they
are not the same product. Muse Glimmer is dense, so all 30B activate, and its
pitch is that it fits on one consumer GPU at 4-bit. Lightning is sparse, so
3B activate, and its pitch is throughput at long context inside a server-side
agent loop. Same headline number, different constraint being solved — a reminder
that "30B" has stopped being a useful comparison unit.

The licence is the other difference, and it is not new to this wiki: OpenMDW-1.1 is the same licence Laguna S 2.1 shipped under on 2026-07-21, where it is recorded as permitting use, modification and redistribution including commercial use (source). Coverage of this release describes it the same way — open for commercial use without material restrictions (source). Two unrelated vendors picking the same non-Apache open-model licence within three weeks is the more interesting fact, and worth watching rather than concluding from.

Conflicting Reports

None on the model's specification. The Switchyard cost claim is reported two ways — NVIDIA's "nearly one-third of Opus 4.8 alone" and VentureBeat's headline "cutting task costs to a third in its own tests" — which are the same claim, and both label it as NVIDIA's internal benchmark (source).

Open Questions

  • Training compute and token budget — undisclosed. The distillation lineage from Nemotron 3 Ultra is stated; nothing else about the training run is.
  • What PinchBench is, and who maintains it. An 86.5% on a suite this repo has no snapshot of is a number with no denominator.
  • Does 1M context hold up? No long-context retrieval measurement surfaced in anything read — the figure is an architectural capacity, not a demonstrated one.
  • The OpenMDW-1.1 terms. Named, characterised, not read.
  • Was the evaluation harness released with the weights? The distinction Eval Harness Configuration exists to record, unresolved here.

Sources

Referenced by

Gemini 3.7 FlashModel RoutingNVIDIA

Sources