--- title: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 ·... description: NVIDIA has released NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, a 30B parameter (3B active) Mixture-of-Experts LLM using a hybrid Mamba-2 + MoE + Attention... canonical: https://daily.dev/posts/nvidia-nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4-hugging-face-cqfpf3znc twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 · Hugging Face | daily.dev og:description: NVIDIA has released NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, a 30B parameter (3B active) Mixture-of-Experts LLM using a hybrid Mamba-2 + MoE + Attention... og:url: https://daily.dev/posts/nvidia-nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4-hugging-face-cqfpf3znc og:image: https://api.daily.dev/og/posts/cQfPf3ZNc.png og:image:alt: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 · Hugging Face og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 · Hugging Face **[Hacker News](https://daily.dev/sources/hn)** · 31 min read · 0 upvotes · 0 comments ## Summary NVIDIA has released NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, a 30B parameter (3B active) Mixture-of-Experts LLM using a hybrid Mamba-2 + MoE + Attention architecture. It supports up to 1M token context length, runs on a single H100 or DGX Spark (GB10), and is quantized using NVFP4 post-training quantization. The model ships with multiple speculative decoding strategies (DSpark, DFlash, MTP) for faster inference. Benchmark highlights include 52.80 on SWE-bench Verified and 75.57 on GPQA Diamond. Deployment is supported via vLLM, TensorRT-LLM, and SGLang with detailed configuration snippets for 1×H100, 8×H100, 1×GB200, and Ampere GPUs. The model is commercially licensed under OpenMDW-1.1 and pre-trained on 20T+ tokens with a post-training data cutoff of May 2026. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What hardware can run NVIDIA Nemotron-3.5-Lightning-30B on a single GPU? Nemotron-3.5-Lightning-30B-A3B-NVFP4 runs on a single DGX Spark (GB10) or a single H100. It also supports GB200 and, via W4A16 kernels, extends coverage to Ampere-class GPUs. The NVFP4 quantization keeps the 30B total parameter model within single-GPU memory budgets while maintaining near-BF16 benchmark parity. _Engineers choosing GPU hardware for self-hosted LLM inference track model-hardware compatibility like this on daily.dev._ ### What speculative decoding methods does Nemotron-3.5-Lightning support and when should I use each? Nemotron-3.5-Lightning ships with three speculative decoding strategies. DSpark is a semi-autoregressive drafter that proposes a whole token block in one forward pass — recommended for DGX Spark and low-concurrency data centre deployments. DFlash uses a lightweight block-diffusion model for one-pass draft generation. MTP (Multi-Token Prediction) trains the network to predict several future tokens per position and is built into the model itself. _Teams optimizing LLM serving throughput find deployment configuration comparisons like this on daily.dev._ ### What is the SWE-bench Verified score for NVIDIA Nemotron-3.5-Lightning-30B-A3B-NVFP4? The NVFP4 checkpoint scores 52.80 on SWE-bench Verified, slightly above the BF16 baseline of 51.56. On GPQA Diamond it scores 75.57 (vs 75.44 BF16), and on MMLU Pro it scores 81.62 (vs 81.94 BF16), showing that quantization introduces minimal accuracy degradation across reasoning and coding benchmarks. _Developers evaluating models for coding agents track benchmark comparisons like these on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 29 comments across hackernews (as of 2026-08-14). **TL;DR:** The community sees Nemotron-3.5 Lightning as a genuinely interesting open-source model with a compelling hybrid architecture and fully open training pipeline, though most agree it trails Qwen 3.6 35B-A3B on benchmarks and note some quality concerns with the NVFP4 quantization. **Sentiment:** 45% positive · 40% mixed · 15% skeptical **The case for** - NVIDIA's hardware-selling incentive gives it a durable, long-term motivation to keep releasing open-weight models. - The fully open training pipeline, recipes, and partial training data make it uniquely reproducible among competitive models. - Runs at ~50–100 tokens/second on consumer Apple Silicon, making it practically accessible without data-center hardware. - The hybrid Mamba-2 + MoE + Attention architecture is a technically interesting direction that is catching up to pure-transformer models. - Nemotron models are perceived as less benchmark-optimized and more generalizable than some competitors. **The pushback** - Benchmark scores trail Qwen 3.6 35B-A3B by roughly 20% across most tasks despite the same 3B active parameters. - The NVFP4 quantization causes a notable SWE-bench score drop compared to the BF16 version. - Reasoning traces show over-thinking behavior, and the model struggles with certain lateral-thinking tasks even after nudging. - The NVFP4 format raises concerns about proprietary lock-in. **By community** - hackernews (mixed): Broadly positive about the open training pipeline and NVIDIA's structural incentive to keep releasing open models, but tempered by clear benchmark gaps versus Qwen and real quality concerns around the NVFP4 quantization. **Hottest debate:** Whether NVIDIA's open-model releases represent genuine community contribution or are primarily a hardware-sales strategy that commoditizes AI models for NVIDIA's own benefit. **Open questions** - Will a Qwen 3.8 35B-A3B MoE follow, and how much will it improve over 3.6? - Is the OpenRouter free endpoint serving the BF16 or NVFP4 variant? - Does the model's over-thinking behavior improve with an explicit reasoning budget or system prompt? - Why does the Codex harness score so much lower than other agentic coding harnesses in the benchmarks? **Highlights** > This links to the nvfp4 version of the model, so they only compare it to the bf16 in this model card. If you're looking for other similar model comparisons they are in the model card of the bf16 version here [1]. While it looks "behind" the qwen equivalent model on most benchmarks, a few personal notes: - nemotron models feel to me a bit less benchmaxxed / "stubborn". That means that they generalise a bit better, or can be tasked to solve similar but not quite identical task types to the training data (something that's hard to do w/ qwen/ds models) - nemotron series are also open training (w/ open training recipes and some training data public) - nvda will have an incentive to continue this kind of releases, even if other parties slowly abandon the open release of models. Whatever other incentives 3rd party labs have (i.e. meta, goog w/ gemma, the chinese labs that IPOd, etc) nvda will always want to sell hardware so their incentive to keep pushing open models is evident and will likely continue "forever". [1] - https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-... > — [NitpickLawyer on hackernews · 3 comments](https://news.ycombinator.com/item?id=49258797) > You might look at this and and be a bit disappointed by the performance against qwen and gemma models - but this is an entirely open source training pipeline, this is quite impressive and I don't think another model this performant exists with fully open source data and recipes alongside the weights. > — [kamranjon on hackernews · 2 comments](https://news.ycombinator.com/item?id=49258981) > Runs fast on my Mac - ~100 tokens/second - but it's a bit of an over-thinker. It sketched out four different SVGs in the reasoning trace before returning a (bad) pelican riding a bicycle: https://tools.simonwillison.net/markdown-svg-renderer#url=ht... I used LM Studio and this model file: https://lmstudio.ai/models/nvidia/nemotron-3.5-lightning I got a pretty solid (and reasonably fast) result for this simply coding agent activity (with simonw/llm-coding-agent): llm code -m lmstudio/nvidia/nemotron-3.5-lightning \ 'how does auth work?' https://gist.github.com/simonw/a8741f79280cedc86bcb8d90edfb4... - took just under a minute. > — [simonw on hackernews · 2 comments](https://news.ycombinator.com/item?id=49261268) > The Nemotron 3 family uses Mamba-2 because that is what we had validated when we started building the family. Changing the SSM backbone is a major revision, not a dot release. At least for us. > — [bcatanzaro on hackernews](https://news.ycombinator.com/item?id=49262066) > That SWEBench drop going from bf16 to fp4 is massive > — [piyh on hackernews](https://news.ycombinator.com/item?id=49260107) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49257947) · 101 points · 29 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#nvidia](https://daily.dev/tags/nvidia), [#vllm](https://daily.dev/tags/vllm), [#mixture-of-experts](https://daily.dev/tags/mixture-of-experts) [View this post on daily.dev](https://daily.dev/posts/nvidia-nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4-hugging-face-cqfpf3znc) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 · Hugging Face","url":"https://daily.dev/posts/nvidia-nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4-hugging-face-cqfpf3znc","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/nvidia-nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4-hugging-face-cqfpf3znc"},"datePublished":"2026-08-11T15:26:15.434Z","dateModified":"2026-08-14T13:55:41.995Z","description":"NVIDIA has released NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, a 30B parameter (3B active) Mixture-of-Experts LLM using a hybrid Mamba-2 + MoE + Attention...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4fe6eb411a28640fe88feba436a28147?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4fe6eb411a28640fe88feba436a28147?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/nvidia-nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4-hugging-face-cqfpf3znc","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,nvidia,vllm,mixture-of-experts","timeRequired":"PT31M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 · Hugging Face"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/nvidia-nvidia-nemotron-3-5-lightning-30b-a3b-nvfp4-hugging-face-cqfpf3znc#faq","mainEntity":[{"@type":"Question","name":"What hardware can run NVIDIA Nemotron-3.5-Lightning-30B on a single GPU?","acceptedAnswer":{"@type":"Answer","text":"Nemotron-3.5-Lightning-30B-A3B-NVFP4 runs on a single DGX Spark (GB10) or a single H100. It also supports GB200 and, via W4A16 kernels, extends coverage to Ampere-class GPUs. The NVFP4 quantization keeps the 30B total parameter model within single-GPU memory budgets while maintaining near-BF16 benchmark parity. Engineers choosing GPU hardware for self-hosted LLM inference track model-hardware compatibility like this on daily.dev."}},{"@type":"Question","name":"What speculative decoding methods does Nemotron-3.5-Lightning support and when should I use each?","acceptedAnswer":{"@type":"Answer","text":"Nemotron-3.5-Lightning ships with three speculative decoding strategies. DSpark is a semi-autoregressive drafter that proposes a whole token block in one forward pass — recommended for DGX Spark and low-concurrency data centre deployments. DFlash uses a lightweight block-diffusion model for one-pass draft generation. MTP (Multi-Token Prediction) trains the network to predict several future tokens per position and is built into the model itself. Teams optimizing LLM serving throughput find deployment configuration comparisons like this on daily.dev."}},{"@type":"Question","name":"What is the SWE-bench Verified score for NVIDIA Nemotron-3.5-Lightning-30B-A3B-NVFP4?","acceptedAnswer":{"@type":"Answer","text":"The NVFP4 checkpoint scores 52.80 on SWE-bench Verified, slightly above the BF16 baseline of 51.56. On GPQA Diamond it scores 75.57 (vs 75.44 BF16), and on MMLU Pro it scores 81.62 (vs 81.94 BF16), showing that quantization introduces minimal accuracy degradation across reasoning and coding benchmarks. Developers evaluating models for coding agents track benchmark comparisons like these on daily.dev."}}]} ```