# Meta Muse Glimmer vs. NVIDIA Nemotron 3.5 Lightning: A Practical Comparison Meta released [Muse Glimmer 30B](https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model) on August 10, 2026. NVIDIA released [Nemotron 3.5 Lightning](https://developer.nvidia.com/blog/nvidia-nemotron-3-5-lightning-delivers-fast-accurate-specialized-task-execution-for-long-running-agents/) the following day. **Both are 30-billion-parameter open-weight models designed to run on a single consumer GPU and targeting always-on local agent workflows**. The architectural choices behind them are quite different, and so are the practical results. This article covers what each model is, what the specs say, and what happens when you actually put them through complex agentic tasks. ## Muse Glimmer Muse Glimmer is a dense causal transformer released by Meta Superintelligence Labs under the Apache 2.0 license. It's the first open-weight model from the division since the Scale AI acquihire brought Alexandr Wang in to lead the group. The model was distilled from Muse Spark, Meta's larger proprietary model, using logit distillation. ![A screenshot of the Meta AI article explaining the training phases of Muse Glimmer, highlighting the use of "logit distillation."](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/2ba4b4dc-75dd-4c94-7b06-c522fcdfce00/lg2x =1920x1080) Logit distillation trains a smaller student model on the raw output score distributions of a larger teacher model, rather than just its final predictions. This gives the student a richer training signal, allowing it to learn more of the teacher's reasoning patterns. The training process combined logit distillation with supervised fine-tuning, on-policy distillation, and reinforcement learning across general, reasoning, coding, and agentic domains. The architecture is a 29.6B dense transformer paired with a 1.8B ViT-G/14 perception encoder for vision, giving it native multimodal capability. Context length is 131,072 tokens. The model is available at [meta-models/Muse-Glimmer-30B](https://huggingface.co/meta-models/Muse-Glimmer-30B) on Hugging Face. ### DFlash speculative decoding Standard language models generate one token at a time. Muse Glimmer includes DFlash, a lightweight block-diffusion drafter model that proposes chunks of tokens simultaneously. The main model verifies these proposals in parallel, accepting correct tokens and correcting wrong ones. ![A side-by-side comparison in a terminal window showing code generation speed with DFlash Speculative Decoding enabled (57.8 tok/s) versus disabled (26.2 tok/s).](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/c19c8341-98de-45d1-b568-ec002dc07f00/public =1920x1080) The speed difference is significant: 57.8 tok/s with DFlash versus 26.2 tok/s without it on the same hardware, with identical output quality. ## Nemotron 3.5 Lightning NVIDIA's Nemotron 3.5 Lightning uses a hybrid Mixture-of-Experts architecture combining interleaved Mamba-2 layers, MoE layers, and select Attention layers. Total parameter count is 30B; active parameters per forward pass are 3B. ![An animated diagram illustrating the Mixture-of-Experts (MoE) architecture, showing that while the model has 30B parameters "on paper," only 3B are "doing work" at any given time.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/065ea09e-d38e-4b14-8bff-c80472226f00/md2x =1920x1080) The sparse activation is what makes Lightning fast. Because only 3B parameters engage per token regardless of total model size, the computational cost per request stays low while the full 30B parameter pool remains available for routing. NVIDIA reports up to 4x the throughput of comparable models at roughly 410 tokens per second. Lightning is distilled from Nemotron 3 Ultra and released under the OpenMDW-1.1 license, which permits commercial use and modification. The weights, training data, and training recipes are all published. It's available on Hugging Face at [nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16), on OpenRouter with a free tier, and as a NIM microservice at [build.nvidia.com](https://build.nvidia.com). ### NeMo Switchyard NVIDIA released NeMo Switchyard alongside Lightning. Switchyard is an open-source routing library that dispatches different steps of an agent workflow to different models: Lightning handles high-volume execution work like tool calls, result validation, and subagent delegation, while a larger reasoning model handles planning and orchestration. ## Specs at a glance | Feature | Muse Glimmer | Nemotron 3.5 Lightning | | :--- | :--- | :--- | | Parameters | 30B dense | 30B MoE (3B active) | | Architecture | Dense transformer + ViT-G/14 | Hybrid Mamba-2 + MoE + Attention | | Vision | Yes | No | | Context window | 131,072 tokens | 1,000,000 tokens | | Tool calling | Yes | Yes | | License | Apache 2.0 | OpenMDW-1.1 | ![A split-screen graphic comparing the key features of Meta Muse Glimmer and NVIDIA Nemotron 3.5 Lightning.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/67b707ab-9aa4-46fd-70af-2d359e81f700/orig =1920x1080) The headline differences: Muse Glimmer has built-in vision and Nemotron doesn't. Nemotron has a 1M token context window, nearly 8x larger. On paper, Nemotron should dominate long-document tasks and Muse Glimmer should lead on anything requiring image understanding. ## Testing environment The tests ran on a local workstation with an RTX 5090 and 64 GB DDR5 RAM. ![A shot of the physical computer workstation used for testing, showing the RTX 5090 GPU and case fans.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/fcd379e3-9f28-469a-def1-9a7c25644200/md2x =1920x1080) Both models loaded via Ollama. Interactions went through OpenCode. Two MCP servers were running: a custom Adobe Premiere plugin for external editing control, and an Epidemic Sound server for searching and downloading music via the Epidemic Sound API. ## Challenge 1: Interactive ESP32 web dashboard The first task was a complex single-prompt web development job: generate a single-page interactive dashboard visualizing electrical current flow for an ESP32-C3 circuit on a breadboard, with an SVG wiring diagram, interactive tooltips, and pin layouts. **Nemotron 3.5 Lightning:** After five minutes of processing, it produced a blank HTML page with non-functional buttons. No diagram, no interactivity, no useful content. A second attempt with a follow-up prompt produced the same result. **Muse Glimmer:** It produced a partially functional webpage with visual elements representing the microcontroller, pin layouts, and working interactive tooltips. The full wiring diagram wasn't generated, but the result was substantially closer to the prompt's requirements. **Result:** Muse Glimmer wins. Nemotron produced nothing usable. ## Challenge 2: Automated video editing with tool calling This tested multi-step agentic reasoning and sequential tool use. The task: 1. Read a video transcript 2. Find every instance of the word "AI" 3. Create a montage sequence of those clips in Adobe Premiere 4. Search Epidemic Sound for a comedic background track 5. Download the track and add it on a separate audio track 6. Trim the music to match the montage and adjust levels **Nemotron 3.5 Lightning:** It identified the transcript and created the clip sequence correctly. Then it placed the downloaded music onto the same audio track as the speaker's voice, overwriting the original audio. When asked to fix this, it became confused and began deleting tracks from the original timeline. **Muse Glimmer:** It completed the entire workflow on the first pass. Montage created correctly, music found and downloaded, placed on a separate track, trimmed with a fade-out. It also successfully handled a follow-up prompt to swap the music for a different track. **Result:** Muse Glimmer wins. It correctly handled the full workflow; Nemotron damaged the source material. ## Challenge 3: Long-context financial document analysis This was designed to test Nemotron's main theoretical advantage. The task: analyze NVIDIA's FY2026 10-K filing from the SEC, extract the percentage of total revenue from the two largest direct customers, identify the business segment, and cross-reference with the Notes to the Financial Statements. **Nemotron 3.5 Lightning:** Its tool call to fetch the document from the SEC website failed. Rather than retry or troubleshoot, it fell back to a generic search and hallucinated an answer: it claimed the top two customers accounted for 23% of revenue (split 13% and 10%). The actual figures in the document are 36% (split 22% and 14%). **Muse Glimmer:** Despite its much smaller context window, it successfully processed the document, extracted the correct figures (22% and 14%), and confirmed they appeared in the Notes to the Financial Statements. **Result:** Muse Glimmer wins. It correctly answered a task designed to favor its competitor's main architectural advantage. ## Final results | Challenge | Muse Glimmer | Nemotron 3.5 Lightning | | :--- | :--- | :--- | | Webpage build | Partial success | Failure | | Premiere edit | Success | Failure (deleted audio) | | 10-K analysis | Success | Failure (hallucinated numbers) | | **Score** | **2/3** | **0/3** | The result is straightforward. Muse Glimmer is the more reliable model for complex agentic tasks in this comparison. Nemotron 3.5 Lightning's failure pattern is significant: it doesn't just produce mediocre results, it actively causes damage (overwriting source audio, providing false financial figures). ## Context for these results These results reflect a specific set of tasks run on specific hardware with a specific setup. They don't capture everything Nemotron was designed for. **NVIDIA's framing for Lightning is that it's an execution-layer model, meant to handle high-volume routine calls in a system where a larger planning model handles orchestration**. Running Lightning as a standalone agent for complex creative and analytical tasks may not be its intended deployment pattern. NeMo Switchyard exists precisely to route task steps to the most appropriate model, with Lightning handling the repetitive calls and a frontier model handling planning. That said, Muse Glimmer is targeting the same local hardware, the same developer audience, and a similar agentic workflow pitch. The comparison is fair in that context. Both models are worth tracking. **Muse Glimmer's benchmark results are strong: MCP Atlas at 75.5 against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B.** NVIDIA's roadmap for the Nemotron family continues with Nemotron 3 Ultra handling frontier-level orchestration while Lightning handles execution throughput. The architecture is interesting; the current execution needs work. For developers choosing between them today for local agentic workflows, Muse Glimmer is the more reliable option.