NVIDIA Nemotron 3.5 Lightning Review 2026: First Open-Source Model — 4x Faster, $0.05/M, Agent-Ready
Verdict first: Nemotron 3.5 Lightning is NVIDIA's first open-weight model, and it is not a frontier chatbot — it is a purpose-built execution-layer worker for AI agents. At ~31.6B total parameters (only ~3.6B active per token), it produces output up to 4x faster than comparable open models for a median $0.05 per million input tokens, fully free to commercialize under OpenMDW-1.1. If you run high-volume agent workloads — tool calls, code review, document parsing — this is the most cost-effective worker model to ship in 2026. Just don't expect it to out-think GPT-5.6 or Fable 5.
NVIDIA Nemotron 3.5 Lightning — NVIDIA's first open-weight LLM, built to be the fast, cheap engine of your AI agents
What Is Nemotron 3.5 Lightning?
Released on August 11, 2026, Nemotron 3.5 Lightning is NVIDIA's first fully open-weight model — a direct follow-through on CEO Jensen Huang's public endorsement of open models. It is a sparse Mixture-of-Experts LLM of about 31.6 billion total parameters with only ~3.6 billion active per token, interleaving Mamba-2 linear-attention, MoE, and selective-attention layers. That hybrid design trades a little raw reasoning depth for exceptional throughput.
NVIDIA's positioning is unusual and honest: this is a worker model, not a thinker. It is meant to sit underneath a frontier model in an agent pipeline — executing tool calls, reviewing code, triaging security alerts, and parsing documents at high volume and rock-bottom cost.
Specifications at a Glance
| Item | Value |
|---|---|
| Vendor | NVIDIA |
| Release date | August 11, 2026 |
| Positioning | Open-source agent execution layer (worker) |
| Architecture | Hybrid MoE: Mamba-2 + MoE + selective attention |
| Total / active params | ~31.6B / ~3.6B active per token |
| Context window | Up to 1M tokens advertised (28.7K–1.05M by provider; ~256K on a single H100) |
| Inputs | Text only (no image/video) |
| License | OpenMDW-1.1 — free to download, modify, commercialize |
| Training | 20+ trillion tokens, agentic RL post-training |
| Runs on | RTX PC, DGX Spark, DGX Station, Jetson, single H100 |
Pricing: How Cheap Is It?
Because the weights are open, NVIDIA sets no API price. The figures below are median third-party hosted rates (USD per million tokens):
| Item | Nemotron 3.5 Lightning | Median comparable model |
|---|---|---|
| Input | $0.05 | $0.18 |
| Output | $0.20 | $0.40 |
| Blended (7:2:1) | $0.07 | — |
Free ways to run it: OpenRouter's :free endpoint, build.nvidia.com trial endpoint, and GMI's BF16 hosting at $0/$0 — or simply self-host, since the weights cost nothing (you only pay for hardware).
The headline cost story: paired with NVIDIA's NeMo Switchyard routing library, a LangChain evaluation of 145 multi-round agent tasks reported a ~74% cost reduction, and agent workloads ran at roughly one-third the cost of running Opus 4.8 alone. These savings are vendor-reported and pending broader third-party verification — treat them as directional.
Benchmarks & Speed
Figures from Artificial Analysis (August 2026), shown as BF16 / NVFP4 (4-bit):
| Benchmark | Score |
|---|---|
| Intelligence Index | 24 (vs median 9 for similar-size open models; below Nemotron 3 Super's 26) |
| MMLU Pro | 81.94 / 81.62 |
| GPQA Diamond | 75.44 / 75.57 |
| SWE-bench Verified | 51.56 / 52.80 |
| GDPval-AA v2 (agentic) | Elo 824 — beats gpt-oss-120b and Nemotron 3 Super |
Speed: median 299.6 tokens/s output (up to 578 t/s measured), ~1.05s median time-to-first-token. On PinchBench it completed 10,000 tasks 30% faster than Qwen 3.6 35B with comparable accuracy. Notably, the 4-bit NVFP4 quantization shows almost no quality loss — a huge plus for cheap self-hosting.
The takeaway: this is mid-tier general intelligence, top-tier agent throughput. Use it where volume and cost matter more than reasoning depth.
Why Did NVIDIA Suddenly Open-Source a Model?
After Huang's public embrace of open models, this is the company's first open-weight release — and the strategy is transparent: the more open models get run, the more NVIDIA GPUs get sold. Lowering the cost of inference actually drives hardware demand. For enterprises, the upside is a big-vendor-backed, commercially free model with no lock-in.
Pros & Cons
Pros
- ✅ Up to 4x faster output than comparable open-weight models
- ✅ Extremely cheap, with free tiers and free self-hosting
- ✅ Fully open weights — free to commercialize, no vendor lock-in
- ✅ Runs on a single GPU, even consumer RTX cards
- ✅ 4-bit quantization with almost no quality loss
- ✅ Strong agent-task performance (beats Nemotron 3 Super on GDPval-AA)
Cons
- ❌ Not a frontier model — weak at complex reasoning and creative work
- ❌ Text-only, no multimodal input
- ❌ Effective context varies by deployment (~256K on a single H100)
- ❌ Hosted pricing varies across providers — shop around
- ❌ Cost-savings claims are vendor-reported, awaiting independent validation
Who Should Use It (And Who Shouldn't)
Great fit
- Developer teams running AI agents: tool calling, code review, document parsing, security triage
- Enterprises cutting inference costs without accepting vendor lock-in
- Local / edge inference enthusiasts (RTX PCs, Jetson, DGX Spark)
- Researchers fine-tuning an open base model
Poor fit
- Users who want the strongest reasoning, writing, or creative quality
- Multimodal applications (image/video understanding)
- As the sole brain of a general-purpose assistant
Early enterprise adopters include CrowdStrike, Harvey, CodeRabbit, and Boomi — all high-volume, cost-sensitive agent workloads.
Frequently Asked Questions
Is NVIDIA Nemotron 3.5 Lightning free?
Yes, the weights are free under the OpenMDW-1.1 license — download, modify, and commercialize without paying NVIDIA. Third-party hosted APIs vary; several offer free tiers (OpenRouter :free, build.nvidia.com trial).
How big is it?
~31.6B total MoE parameters, with only ~3.6B active per token — an efficient sparse architecture that also fits on a single GPU.
Is it really 4x faster?
Independent measurements show a median ~300 tokens/s output (up to 578), roughly 4x faster than comparable open models, with ~30% faster task completion on PinchBench.
How much does it cost?
Median hosted prices are ~$0.05/M input and ~$0.20/M output. Self-hosting costs only hardware; NVIDIA reports ~74% cost reduction on agent workloads when paired with NeMo Switchyard.
Can I run it locally?
Yes — single GPU (RTX PC, DGX Spark, DGX Station, Jetson, H100). NVFP4 4-bit quantization barely degrades quality.
Is it multimodal?
No, text-only. Pair it with a frontier multimodal model for planning and perception.
What is it best for?
The execution layer of AI agents: high-volume tool calls, code review, security triage, and document parsing. Not a general assistant brain.
Sources
- Artificial Analysis — Nemotron 3.5 Lightning: Intelligence, Performance & Price Analysis
- NVIDIA Nemotron 3.5 Lightning: Specs & Benchmarks (August 2026)
- Nemotron 3.5 Lightning vs Frontier Models: Scorecard
- Chinese tech media coverage: NVIDIA's latest open-source model goes live (Dahe Caifu / 大河财立方)
没有评论:
发表评论