Trending on Kingy
Keep reading with the stories getting the most attention now.
THE ANSWER IN ONE SENTENCE. NVIDIA Nemotron 3.5 Lightning is one of the most practical open models yet for high-volume, text-only agent execution: fast, cheap, customizable and capable of million-token context, while falling short of the strongest 30B-class reasoners and retaining a 30B-class memory footprint.
1. Quick verdict
Worth using? Yes, if you give it the right job. Nemotron 3.5 Lightning is best treated as an execution engine beneath a smarter orchestrator: the model that formats results, calls tools, validates routine outputs and processes long streams of repetitive work. NVIDIA’s own numbers show why. Lightning sits dramatically farther right on the speed curve than other sub-40B open models, and its launch-day OpenRouter price is only $0.10 per million input tokens and $0.25 per million output tokens.
It is a less convincing choice as the one model that must do everything. In NVIDIA’s controlled evaluation, Qwen3.6 35B-A3B is materially stronger on hard science, software engineering, terminal work and long-context accuracy. Meta’s day-old Muse Glimmer 30B adds image input, broader language coverage and a polished 24/32 GB local deployment story. Lightning wins the narrow contest NVIDIA designed it for: cheap, fast, persistent text execution.
Evidence scope: Kingy reviewed NVIDIA’s launch post and model cards, Meta’s Muse Glimmer model card, current Hugging Face port metadata, OpenMDW licensing materials and OpenRouter’s live model catalog on August 11, 2026. Kingy did not run the model, reproduce throughput, or independently verify any benchmark.
| BEST FOR | SKIP IF |
|---|---|
| Text-only agent workers that make many routine tool calls | You need image or screenshot understanding |
| Cost-sensitive routing, extraction, validation and RAG | You want the strongest 30B-class coding or reasoning model |
| Teams that plan to fine-tune a narrow execution policy | Your stack is unprepared for a new hybrid Mamba/MoE architecture |
| Long-running jobs where throughput matters more than peak intelligence | You read “3B active” as “fits like a 3B model” |
| STRENGTHS | WEAKNESSES |
|---|---|
| 30B total parameters, only 3B active per token | All 30B weights still have to live somewhere |
| Up to 1M-token context and controllable reasoning | Long-context accuracy trails key peers in NVIDIA’s own table |
| Very low launch-day API price | Text-only; six named natural languages, with English primary |
| Open weights, training data and recipes under OpenMDW-1.1 | Best-optimized path leans heavily into NVIDIA hardware/software |
| MTP, DSpark and DFlash speculative-decoding options | Launch-day backends and quantizations are still moving targets |
2. What “30B with 3B active” actually means
Nemotron 3.5 Lightning is a hybrid model: interleaved Mamba-2, mixture-of-experts (MoE) and selected attention layers. It contains roughly 30 billion parameters in total, but its router selects only a fraction of its experts for each token, leaving about 3 billion parameters active at a time. That is the source of the speed advantage.
The easy mistake is to turn an execution figure into a memory figure. Only activating 3B parameters does not shrink the stored model to 3B. The unused experts still need to be available when the router selects them. Quantization can compress the weights, but sparsity alone does not remove them.
The user-supplied MLX BF16 port makes the distinction concrete: it is unquantized and listed at 63.2 GB. MLX Community now also lists a 4-bit port at 17.8 GB, which is far more realistic for Apple Silicon. The operating system, KV cache and runtime still compete for the remaining unified memory.
So the shorthand is:
- 30B total describes model capacity and stored weights.
- 3B active describes approximate compute per generated token.
- Quantization determines how much memory the weights consume.
- Context length and concurrency determine how much additional cache headroom the deployment needs.
3. How this analysis was sourced
Kingy reviewed the launch materials without running a hands-on benchmark. The source hierarchy was deliberately narrow:
| EVIDENCE TYPE | WHAT WE USED | LIMIT |
|---|---|---|
| Primary launch material | NVIDIA launch post | Vendor positioning and vendor-selected examples |
| Technical model details | NVIDIA BF16 model card | NVIDIA’s own evaluation setup |
| Primary comparator | Meta Muse Glimmer 30B model card | Meta’s own evaluation setup; not directly interchangeable with NVIDIA’s |
| Availability and footprint | Hugging Face model metadata and MLX port pages | Community ports can change quickly after launch |
| Live pricing | OpenRouter public model catalog | Launch-day list prices can change by provider |
| License interpretation | OpenMDW official FAQ | Not legal advice |
No Kingy latency, accuracy, energy, memory or long-duration reliability measurements are presented here. Vendor benchmark values remain vendor benchmark values even when the evaluation recipes are published.
4. Specifications and costs
The headline specification is attractive, but the deployment footnotes matter just as much.
| SPECIFICATION | NEMOTRON 3.5 LIGHTNING | PRACTICAL LIMIT |
|---|---|---|
| Architecture | 30B-total / 3B-active hybrid Mamba-2 + MoE + attention | Sparse compute with a 30B-class storage footprint |
| Input/output | Text in, text out | No native image, audio or video input |
| Context | Up to 1M tokens | NVIDIA’s single-H100 BF16 recipe uses 256K; full 1M examples use 8× H100 or 1× GB200 |
| Languages | English and coding languages; Spanish, French, German, Italian and Japanese | English is primary; not the 100+ language story Meta claims for Muse Glimmer |
| Precision releases | BF16 reference; NVFP4 deployment release; W4A16 and GGUF paths | Backend and hardware support vary |
| Reasoning | On/off plus runtime thinking budget | More thinking increases output tokens, latency and cost |
| Fine-tuning | SFT, RL, distillation and domain adaptation | Requires a real evaluation set; generic benchmark gains do not guarantee policy accuracy |
| License | OpenMDW-1.1; commercial use permitted | Redistributions must preserve the license and notices; litigation-termination terms apply |
API cost: cheap enough to change routing decisions
At launch, OpenRouter lists Nemotron 3.5 Lightning at $0.10/M input and $0.25/M output, with cache reads at $0.05/M. A free route is also listed, but treat it as an evaluation convenience rather than a production availability promise.
For a simple monthly workload of 10 million uncached input tokens and 2 million output tokens, the launch-day list-price comparison is:
| MODEL | INPUT / OUTPUT PER 1M | ILLUSTRATIVE MONTHLY COST | WHAT YOU ARE PAYING FOR |
|---|---|---|---|
| Nemotron 3.5 Lightning | $0.10 / $0.25 | $1.50 | Fast text execution, long context |
| Meta Muse Glimmer 30B | $0.35 / $1.50 | $6.50 | Dense multimodal agent and stronger all-round local story |
| Qwen3.6 35B-A3B | $0.15 / $1.00 | $3.50 | Better hard-reasoning/coding results in NVIDIA’s table |
| Gemma 4 26B-A4B | $0.12 / $0.40 | $2.00 | Strong instruction following and broader generalist appeal |
| GPT-OSS 20B | $0.03 / $0.13 | $0.56 | Lowest raw API cost, with lower agentic scores in NVIDIA’s suite |
That example excludes caching, provider fees, retries and reasoning-token expansion. It is useful as a relative comparison, not a forecast.
Self-hosting cost: weights are free; infrastructure is not
NVIDIA’s BF16 card supports a single H100 80 GB or A100 80 GB deployment. Baseten’s public dedicated-instance table currently lists those GPUs at roughly $6.50 and $4.00 per hour, respectively, before engineering, storage, observability and idle-capacity decisions. The optimized NVFP4 release is the sensible deployment path for most teams; the BF16 release is explicitly positioned as the customization and research checkpoint.
Local inference moves the bill from tokens to hardware and electricity. A 17.8 GB 4-bit MLX weight set may load on a 24 GB-class machine, but useful context and system headroom can turn “loads” into a very different question from “runs comfortably.” Kingy has not tested the Apple, RTX, DGX Spark or CPU/GGUF paths.
5. Nemotron vs Meta Muse Glimmer 30B
Meta’s Muse Glimmer 30B, released one day earlier, is the comparison that matters because the two models target local, long-running agents at almost the same total parameter count. They get there through opposite design choices.
| QUESTION | NEMOTRON 3.5 LIGHTNING | META MUSE GLIMMER 30B |
|---|---|---|
| Core design | Sparse 30B MoE, 3B active | Dense 29.6B transformer including a ~1.8B vision encoder |
| Modalities | Text → text | Text + image → text |
| Context | Up to 1M | 131,072+ |
| Local deployment | NVIDIA NVFP4/GGUF/MLX ecosystem; BF16 card targets 80 GB GPUs | Meta publishes 4-bit variants for 24/32 GB hardware |
| Speculative decoding | MTP plus DSpark and DFlash options | DFlash drafter; Meta reports 233.4 tok/s on RTX 5090 |
| Languages | Six named natural languages; English primary | Trained on more than 100 languages |
| License | OpenMDW-1.1 | Apache 2.0 |
| OpenRouter price | $0.10 input / $0.25 output | $0.35 input / $1.50 output |
| Best reason to choose it | Throughput, cost, long text context and specialization | Multimodal local agent, broader general capability and simpler consumer positioning |
Where Nemotron is stronger
It is dramatically cheaper to call. On launch-day OpenRouter pricing, Muse Glimmer’s input is 3.5 times as expensive and its output is six times as expensive. In a long-running agent, output-heavy retries and tool narration can make that gap more important than the prompt price.
It has the cleaner role in a routed system. NVIDIA does not pretend Lightning should solve every problem. NeMo Switchyard is designed to send hard planning to a frontier model and routine execution down to Lightning. That division of labour is coherent, measurable and easy to budget.
Its maximum context is far larger. One million tokens creates room for repositories, long records and persistent agent state that exceed Muse Glimmer’s native 131K window. The caveat is crucial: fitting a window is not the same as using it accurately, and the full window can demand much larger hardware.
It is unusually customizable. NVIDIA is releasing weights, training data, recipes and an agentic terminal RL dataset, with documented SFT and reinforcement-learning paths. For teams teaching one repeatable policy such as triage, classification, tool routing or compliance formatting, that may matter more than zero-shot benchmark rank.
Where Muse Glimmer is stronger
It can see. Muse Glimmer accepts screenshots, charts, documents and interleaved images. That immediately makes it the more natural fit for desktop agents, visual QA, chart interpretation and document workflows.
Its consumer-hardware story is more explicit. Meta publishes two quantized targets: one for 32 GB hardware with a reported 0.2% average degradation across 15 benchmarks, and a 17 GB variant for 24 GB hardware with a reported 1.0% degradation. NVIDIA has broad local tooling, but the full-precision MLX port linked in the brief is 63.2 GB and the safest hardware guidance still depends heavily on quantization and context.
Meta reports stronger all-round agent results. Muse Glimmer’s card reports 76.0 on SWE-bench Verified, 51.7 on TerminalBench 2.1, 75.5 on MCP Atlas and 74.6 on DeepSearch QA. Those are impressive figures, but they are not a clean head-to-head with NVIDIA’s numbers: scaffolds, prompts, reasoning settings and evaluator versions can move results materially.
Choose Muse Glimmer as the local agent you interact with. Choose Nemotron as the worker your agent calls repeatedly.
6. Strengths and weaknesses in NVIDIA’s own benchmark table
The most useful benchmark evidence is NVIDIA’s controlled comparison, because it evaluates Lightning, Qwen3.6 35B-A3B, Gemma 4 26B-A4B and GPT-OSS 20B under one set of NVIDIA recipes.
| BENCHMARK | LIGHTNING | QWEN3.6 35B | GEMMA 4 26B | GPT-OSS 20B |
|---|---|---|---|---|
| IFBench, loose ↑ | 71.88 | 63.71 | 77.25 | 68.50 |
| PinchBench ↑ | 85.37 | 88.07 | 74.70 | 57.20 |
| BrowseComp ↑ | 36.97 | 48.74 | 26.30 | n/a |
| GPQA Diamond ↑ | 75.44 | 83.40 | 79.61 | 71.46 |
| SWE-bench Verified ↑ | 51.56 | 70.12 | 57.40 | 52.44 |
| AA-LCR long context ↑ | 52.00 | 61.06 | 57.56 | 32.88 |
Three conclusions stand out.
First, Lightning is not the quality leader. Qwen3.6 35B-A3B beats it across hard reasoning, coding, browsing and long-context accuracy. Gemma also beats it on GPQA, SWE-bench and AA-LCR.
Second, Lightning’s best result is role-specific. It comes close to Qwen on PinchBench while NVIDIA says it completes 10,000 tasks 30% faster at comparable accuracy. That is exactly what an execution-layer model should optimize.
Third, the 1M context headline needs restraint. Lightning supports the larger window, but scores 52.0 on NVIDIA’s long-context measure versus 61.06 for Qwen and 57.56 for Gemma. Capacity is not comprehension.
7. The feature that matters most: model routing
Nemotron 3.5 Lightning makes the most sense when it is not the top of the stack.
Give planning, ambiguous judgment and difficult code repair to a stronger model. Give Lightning the repeated execution steps: normalize this response, call this function, verify these fields, compare the result to policy, extract the next action, repeat. NVIDIA’s new NeMo Switchyard library formalizes that pattern by routing tasks between open and closed models.
That is also how to evaluate Lightning internally. Do not begin with a grab-bag benchmark and ask whether it is “smart.” Begin with a frozen set of the exact steps consuming your agent budget:
- Route identical tasks through the current model and Lightning.
- Score exact schema validity, tool selection, retry rate and final task success.
- Measure complete-task latency and tokens instead of generation speed alone.
- Track escalations to the frontier model and false confidence separately.
- Fine-tune only after the failure taxonomy is stable.
NVIDIA highlights a CodeRabbit routing experiment that reportedly cost under $100 to post-train and improved exact route agreement over a GPT-class baseline. It is an encouraging example of the intended use, not proof that every specialist task will transfer as cleanly.
8. Deployment and compatibility realities
| TARGET | PRACTICAL PATH | WHAT TO CHECK FIRST |
|---|---|---|
| NVIDIA data centre | NVFP4 with TensorRT-LLM, vLLM or SGLang | GPU architecture, concurrency and supported cache mode |
| H100/A100 research | BF16 on one 80 GB GPU, usually 256K context | Nightly vLLM requirements and Mamba cache memory |
| Full 1M BF16 context | 8× H100 with expert parallelism or 1× GB200 example | Cost, cache headroom and real retrieval quality |
| DGX Spark | NVFP4 DSpark recipe | Drafter compatibility and low-concurrency workload shape |
| GeForce / local PC | Official/community GGUF or optimized quantization | VRAM, llama.cpp support and tool-call template fidelity |
| Apple Silicon | MLX 4-bit/8-bit community ports | Unified memory after weights, cache and macOS overhead |
| Hosted API | OpenRouter or launch partners | Actual provider context, rate limits, privacy and uptime |
NVIDIA lists support across Blackwell, Hopper and Ampere, and names LM Studio, llama.cpp, Ollama and Unsloth in the launch ecosystem. That breadth is useful, but “listed” is not the same as “mature at every quantization.” Hybrid Mamba/MoE models often need backend-specific kernels and parser support. NVIDIA’s own quick start points to a nightly vLLM image and uses specific nemotron_v3 reasoning and qwen3_coder tool parsers.
For Mac users, the linked BF16 MLX conversion is best understood as a reference port for high-memory machines. The 17.8 GB 4-bit MLX conversion is the practical starting point. For PC users, GGUF will be attractive, but confirm that the selected runtime handles the architecture, chat template, reasoning output and tool calls correctly. Loading the file is the first check, not the last.
9. Who should use which model?
Choose Nemotron 3.5 Lightning for high-volume text execution, cheap tool calling, long-context extraction, routing and narrow post-training. It is especially compelling when a stronger model already handles planning.
Choose Meta Muse Glimmer 30B for a local generalist that must understand images, screenshots and documents, or when 24/32 GB consumer deployment is the central requirement.
Choose Qwen3.6 35B-A3B when the job is still text-only but hard reasoning, coding and terminal performance matter more than Lightning’s throughput and token price.
Choose Gemma 4 26B-A4B when instruction following and general-purpose quality matter more than million-token context, especially if its ecosystem already fits your stack.
Choose GPT-OSS 20B when the lowest hosted token cost is the dominant constraint and your own task evaluation shows its lower benchmark ceiling is acceptable.
10. Nemotron 3.5 Lightning FAQ
Is Nemotron 3.5 Lightning really a 3B model?
No. It is a roughly 30B-total MoE model that activates about 3B parameters per token. Its compute can resemble a much smaller model, but its weights still reflect the full model.
Can it run on a Mac?
Yes, through community MLX ports. The BF16 port linked in the brief is 63.2 GB. A 4-bit MLX port is listed at 17.8 GB, making it much more realistic, though usable context and speed depend on the machine and remaining unified memory. Kingy has not tested either port.
Is it better than Meta Muse Glimmer 30B?
For cheap text throughput and very long context, likely yes. For multimodal local agents and a broader one-model setup, Muse Glimmer is the stronger fit. There is no independent, same-setup head-to-head yet.
Does it really support one million tokens?
NVIDIA lists up to 1M tokens. Its published BF16 example uses 8× H100 or 1× GB200 for the full window, while the single-H100 recipe uses 256K. More context also does not guarantee better retrieval; Lightning trails Qwen and Gemma on NVIDIA’s AA-LCR evaluation.
How much does it cost?
The weights are free to use under OpenMDW-1.1. OpenRouter listed the model at $0.10/M input and $0.25/M output on August 11, 2026, plus a free route. Self-hosting adds GPU, storage, operations and engineering costs.
Can it understand images?
No. The official model is text-in/text-out. Muse Glimmer, Gemma-class multimodal models or a separate vision encoder are better choices when screenshots and documents are part of the task.
Can businesses fine-tune and deploy it commercially?
Yes. NVIDIA says the model is ready for commercial use. OpenMDW-1.1 permits use, modification and redistribution with notice-preservation requirements and a patent/copyright litigation-termination clause. This article is not legal advice.
Should I use the BF16 or NVFP4 release?
Use BF16 as the reference for research, evaluation and customization. NVIDIA explicitly recommends NVFP4 for optimized production inference. Community 4-bit MLX and GGUF ports are practical local options, but validate quality and tool calling on your own tasks.
Final verdict: excellent at one layer of the stack
Nemotron 3.5 Lightning is easy to underestimate because it does not top NVIDIA’s own quality table. That misses the point. The model is designed to make competent routine agent work cheap enough and fast enough to run constantly, while stronger models handle the decisions that deserve them.
Its advantages are real: sparse 3B-active compute, a one-million-token ceiling, controllable reasoning, multiple speculative-decoding paths, unusually open customization materials and an API price that undercuts its closest peers. Its constraints are equally real: it is text-only, still stores a 30B-class weight set, leans toward NVIDIA’s optimized stack and loses to Qwen3.6 35B on many difficult tasks.
Meta’s Muse Glimmer 30B is the better single local assistant. Nemotron 3.5 Lightning is the better specialist worker.
Put it behind a router, measure completed work rather than tokens per second, and promote difficult cases to a stronger model. In that role, Nemotron 3.5 Lightning earns a strong recommendation.
Official sources
- NVIDIA Nemotron 3.5 Lightning launch post
- NVIDIA Nemotron 3.5 Lightning BF16 model card
- NVIDIA Nemotron 3.5 Lightning NVFP4 model card
- NVIDIA NeMo Gym reproducibility recipes
- MLX Community BF16 conversion
- MLX Community 4-bit conversion
- Meta Muse Glimmer 30B model card
- OpenMDW-1.1 official FAQ
- OpenRouter live model catalog
- Baseten public cloud pricing
Editorial note: specifications, availability and prices were checked August 11, 2026 and can change. All benchmark and throughput figures are vendor- or platform-published unless explicitly stated otherwise.