🚨 @nvidiaai 𝗝𝗨𝗦𝗧 𝗣𝗨𝗦𝗛𝗘𝗗 𝗧𝗛𝗘 𝗢𝗡𝗘-𝗦𝗣𝗔𝗥𝗞 𝗙𝗥𝗢𝗡𝗧𝗜𝗘𝗥 𝗙𝗢𝗥𝗪𝗔𝗥𝗗! 🇺🇸🔥 🧠 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟: 1 MILLION native context tokens + 84.17% SixCat! ⚡ 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗕𝗢𝗟𝗧 + 𝗟𝗶𝗟𝗶𝗖𝗼𝗿𝗿 𝗗𝗙𝗹𝗮𝘀𝗵: 87.8 tok/s single-stream decode + 195.1 tok/s balanced throughput! 🖥️ 𝗧𝗪𝗢 𝗠𝗢𝗗𝗘𝗟𝗦. 𝗧𝗪𝗢 𝗗𝗜𝗙𝗙𝗘𝗥𝗘𝗡𝗧 𝗦𝗧𝗥𝗘𝗡𝗚𝗧𝗛𝗦. 𝗢𝗡𝗘 𝗗𝗚𝗫 𝗦𝗣𝗔𝗥𝗞! And YES, these are my actual measurements from the DGX Spark sitting in my home! 🤯 After NVIDIA gave me pre-release access to Nemotron 3.5 Super VL, I've spent the past week benchmarking, optimizing, and experimenting with these new models. But I wanted to answer something more important than which model has the highest intelligence score... 𝗪𝗛𝗜𝗖𝗛 𝗠𝗢𝗗𝗘𝗟 𝗔𝗖𝗧𝗨𝗔𝗟𝗟𝗬 𝗚𝗘𝗧𝗦 𝗪𝗢𝗥𝗞 𝗗𝗢𝗡𝗘 𝗙𝗔𝗦𝗧𝗘𝗥, 𝗔𝗡𝗗 𝗪𝗛𝗜𝗖𝗛 𝗢𝗡𝗘 𝗖𝗔𝗡 𝗛𝗔𝗡𝗗𝗟𝗘 𝗠𝗢𝗥𝗘 𝗪𝗢𝗥𝗞? I tested both models separately on the SAME GB10 DGX Spark. Not two Sparks. Not a datacenter cluster. 𝗢𝗡𝗘. 𝗗𝗚𝗫. 𝗦𝗣𝗔𝗥𝗞. And these results are INCREDIBLY interesting for the future of local AI! ━━━━━━━━━━━━━━━━━━━━ 🧠 𝗣𝗔𝗥𝗧 𝟭: 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 — 𝗧𝗛𝗘 𝟭𝗠 𝗖𝗢𝗡𝗧𝗘𝗫𝗧 𝗣𝗢𝗪𝗘𝗥𝗛𝗢𝗨𝗦𝗘! 120B total parameters. 12B active parameters. Multimodal reasoning, coding, tools, image, document and video understanding. And here's what REALLY caught my attention... 𝗔 𝗙𝗨𝗟𝗟 𝟭,𝟬𝟰𝟴,𝟱𝟳𝟲 𝗡𝗔𝗧𝗜𝗩𝗘 𝗖𝗢𝗡𝗧𝗘𝗫𝗧 𝗪𝗜𝗡𝗗𝗢𝗪 𝗢𝗡 𝗢𝗡𝗘 𝗗𝗚𝗫 𝗦𝗣𝗔𝗥𝗞! 🤯 And YES, I've personally tested the FULL 1M-token context window on my single Spark and confirmed it WORKS! 🔥 For comparison, Qwen3.8-Flash-Next has a native context window of 262,144 tokens. Qwen recommends using RoPE scaling techniques such as YaRN to extend beyond its native context. 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 𝗦𝗨𝗣𝗣𝗢𝗥𝗧𝗦 𝟰× 𝗧𝗛𝗔𝗧 𝗡𝗔𝗧𝗜𝗩𝗘 𝗖𝗢𝗡𝗧𝗘𝗫𝗧 𝗪𝗜𝗧𝗛𝗢𝗨𝗧 𝗬𝗮𝗥𝗡! And let me emphasize... 𝗢𝗡𝗘 𝗦𝗣𝗔𝗥𝗞! 𝗡𝗢𝗧 𝗧𝗪𝗢! Why am I so excited about this? Because for everyday and enterprise AI workloads, a massive context window can be MORE useful than simply having a slightly higher intelligence score. Think about the possibilities: ✅ Entire codebases and large repositories ✅ Massive infrastructure logs and incident histories ✅ Enterprise documentation and policies ✅ Long-running autonomous agents ✅ Multimodal document analysis ✅ Extended research and investigation workflows ✅ Larger working histories without constant summarization Instead of repeatedly compressing information to fit inside a smaller context window, you have significantly more room to work! Now, 4× the context capacity doesn't mean 4× the reasoning quality or processing speed. But having a FULL native 1M-token window working on a single Spark that I own at home is a HUGE DEAL! 𝗧𝗛𝗜𝗦 𝗜𝗦 𝗧𝗛𝗘 𝗞𝗜𝗡𝗗 𝗢𝗙 𝗟𝗢𝗖𝗔𝗟 𝗔𝗜 𝗣𝗥𝗢𝗚𝗥𝗘𝗦𝗦 𝗜 𝗪𝗔𝗡𝗧 𝗧𝗢 𝗦𝗘𝗘! 🔥 ━━━━━━━━━━━━━━━━━━━━ ⚡ 𝗣𝗔𝗥𝗧 𝟮: 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗕𝗢𝗟𝗧 — 𝗧𝗛𝗘 𝗟𝗢𝗖𝗔𝗟 𝗦𝗣𝗘𝗘𝗗 𝗠𝗔𝗦𝗧𝗘𝗥! Now let's talk about Bolt, because NVIDIA has been COOKING with this one too! 👀 74B total parameters. Only 9B active. An official 24GB-class IQ2 checkpoint. 262,144-token context support. Designed to bring fast, capable agentic AI to smaller local hardware footprints. But NVIDIA didn't stop at the model itself... They also released 𝗟𝗶𝗟𝗶𝗖𝗼𝗿𝗿, a speculative drafting model used with their 𝗗𝗙𝗹𝗮𝘀𝗵 speculative decoding system. Naturally, I HAD to test it on my Spark! I ran the same Bolt IQ2 target through SixCat Speed with and without the LiLiCorr DFlash drafter. Same GB10 hardware. Same target checkpoint. Same vLLM engine. Same benchmark workloads. And the results were WILD! 🤯 🔥 𝗕𝗢𝗟𝗧 𝗦𝗣𝗘𝗘𝗗: 𝗡𝗢 𝗗𝗥𝗔𝗙𝗧 𝘃𝘀 𝗟𝗜𝗟𝗜𝗖𝗢𝗥𝗥 𝗗𝗙𝗟𝗔𝗦𝗛 ⚡ 𝗦𝗜𝗡𝗚𝗟𝗘-𝗦𝗧𝗥𝗘𝗔𝗠 𝗗𝗘𝗖𝗢𝗗𝗘: 30.9 → 87.8 TOK/S 🚀 𝟮.𝟴𝟰× 𝗙𝗔𝗦𝗧𝗘𝗥! 🔥 𝗕𝗔𝗟𝗔𝗡𝗖𝗘𝗗 𝗧𝗛𝗥𝗢𝗨𝗚𝗛𝗣𝗨𝗧 (𝟴 𝗖𝗢𝗡𝗖𝗨𝗥𝗥𝗘𝗡𝗧 𝗥𝗘𝗤𝗨𝗘𝗦𝗧𝗦): 136.5 → 195.1 OUTPUT TOK/S 🚀 +𝟰𝟯% 𝗛𝗜𝗚𝗛𝗘𝗥 𝗧𝗛𝗥𝗢𝗨𝗚𝗛𝗣𝗨𝗧! 📥 𝗦𝗜𝗡𝗚𝗟𝗘-𝗦𝗧𝗥𝗘𝗔𝗠 𝗣𝗥𝗘𝗙𝗜𝗟𝗟: 1,584 → 1,328 tok/s (−16%) ⏱️ 𝗕𝗔𝗟𝗔𝗡𝗖𝗘𝗗 𝗧𝗜𝗠𝗘 𝗧𝗢 𝗙𝗜𝗥𝗦𝗧 𝗧𝗢𝗞𝗘𝗡: 1.68 → 1.87 seconds (+0.19 seconds) And here's the best part... ✅ All THREE DFlash speed profiles completed 32/32 requests each! ✅ Approximately 55.5% of drafted tokens were accepted! ✅ vLLM speculative decoding counters confirmed LiLiCorr was actively being used! ✅ No memory guard stops or swap growth observed! 𝗔𝗟𝗠𝗢𝗦𝗧 𝟯× 𝗧𝗛𝗘 𝗦𝗜𝗡𝗚𝗟𝗘-𝗦𝗧𝗥𝗘𝗔𝗠 𝗗𝗘𝗖𝗢𝗗𝗘 𝗦𝗣𝗘𝗘𝗗 𝗨𝗦𝗜𝗡𝗚 𝗧𝗛𝗘 𝗦𝗔𝗠𝗘 𝗕𝗢𝗟𝗧 𝗧𝗔𝗥𝗚𝗘𝗧 𝗢𝗡 𝗧𝗛𝗘 𝗦𝗔𝗠𝗘 𝗦𝗣𝗔𝗥𝗞! There are some prefill and latency trade-offs, but the decoding improvement is absolutely impressive. And remember, this is a model with a target checkpoint of approximately 24GB! The 24GB figure is the target-model directory size, not the complete memory footprint with LiLiCorr loaded. ━━━━━━━━━━━━━━━━━━━━ 🧪 𝗣𝗔𝗥𝗧 𝟯: 𝗠𝗬 𝗦𝗜𝗫𝗖𝗔𝗧 𝗤𝗨𝗔𝗟𝗜𝗧𝗬 𝗕𝗘𝗡𝗖𝗛𝗠𝗔𝗥𝗞𝗦! Speed is great, but what about INTELLIGENCE? Both models completed my SixCat evaluation suite: 120 questions. 6 categories. Strict scoring. Thinking enabled. Code executed and graded. 🟢 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 (𝟭𝟮𝟬𝗕-𝗔𝟭𝟮𝗕) 🏆 𝟴𝟰.𝟭𝟳% — 𝟭𝟬𝟭/𝟭𝟮𝟬 𝗣𝗔𝗦𝗦𝗘𝗗! 🟣 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗕𝗢𝗟𝗧 (𝟳𝟰𝗕-𝗔𝟵𝗕) 🏆 𝟳𝟳.𝟱𝟬% — 𝟵𝟯/𝟭𝟮𝟬 𝗣𝗔𝗦𝗦𝗘𝗗! 📊 𝗤𝗨𝗔𝗟𝗜𝗧𝗬 𝗕𝗥𝗘𝗔𝗞𝗗𝗢𝗪𝗡: 𝗕𝗢𝗟𝗧 → 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 🧠 Knowledge: 60% → 85% 🔢 Math: 100% → 100% 🎯 Truthfulness: 80% → 85% 📋 Instruction Following: 50% → 50% 💻 Code: 85% → 100% 🔧 Tools: 90% → 85% 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 𝗪𝗜𝗡𝗦 𝗢𝗩𝗘𝗥𝗔𝗟𝗟 𝗕𝗬 𝟲.𝟲𝟳 𝗣𝗘𝗥𝗖𝗘𝗡𝗧𝗔𝗚𝗘 𝗣𝗢𝗜𝗡𝗧𝗦! But don't underestimate Bolt! ✅ 100% in Math! ✅ 90% in Tool Use! ✅ 77.5% overall with a much smaller model footprint! These scores are from the earlier SixCat quality runs. The LiLiCorr experiment is a new SPEED benchmark, not a new quality evaluation. Both quality receipts also flagged instruction truncation or looping, so the results should be interpreted with those limitations. ━━━━━━━━━━━━━━━━━━━━ 🏎️ 𝗣𝗔𝗥𝗧 𝟰: 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 𝘃𝘀 𝗕𝗢𝗟𝗧 + 𝗗𝗙𝗟𝗔𝗦𝗛 — 𝗧𝗛𝗘 𝗥𝗘𝗔𝗟 𝗧𝗥𝗔𝗗𝗘-𝗢𝗙𝗙! Now that we have both quality and speed measurements, here's how my tested configurations compare. 🟢 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 — 𝗧𝗛𝗘 𝗛𝗜𝗚𝗛-𝗖𝗔𝗣𝗔𝗕𝗜𝗟𝗜𝗧𝗬 𝗟𝗘𝗔𝗗 𝗠𝗢𝗗𝗘𝗟 🏆 SixCat: 84.17% 🧠 Native context: 1,048,576 tokens 📦 Model directory: 75.06 GB ⚡ Single-stream decode: 37.3 tok/s 🚀 Balanced throughput C=8: 90.9 tok/s ⏱️ Long-prompt TTFT p50, C=8: 4.47 seconds Inference: NVIDIA NVFP4, vLLM 0.31.0, FlashInfer, MTP-5. 🟣 𝗕𝗢𝗟𝗧 + 𝗟𝗜𝗟𝗜𝗖𝗢𝗥𝗥 𝗗𝗙𝗟𝗔𝗦𝗛 — 𝗧𝗛𝗘 𝗙𝗔𝗦𝗧 𝗪𝗢𝗥𝗞𝗘𝗥 🏆 SixCat: 77.50% (earlier no-draft quality run) 🧠 Native context: 262,144 tokens 📦 Bolt target directory: 24.12 GB ⚡ Single-stream decode: 87.8 tok/s 🚀 Balanced throughput C=8: 195.1 tok/s ⏱️ Balanced TTFT p50: 1.87 seconds Inference: Mixed ModelOpt IQ2, vLLM 0.32.0rc1 development build, B12X 1.5.1, FlashInfer, LiLiCorr DFlash. 𝗕𝗢𝗟𝗧 + 𝗗𝗙𝗟𝗔𝗦𝗛 𝗗𝗘𝗟𝗜𝗩𝗘𝗥𝗦: 🔥 𝟮.𝟯𝟱× 𝗙𝗔𝗦𝗧𝗘𝗥 than Super VL on measured single-stream decode! 🔥 𝟮.𝟭𝟱× 𝗛𝗜𝗚𝗛𝗘𝗥 balanced output throughput at 8 concurrent requests! But Super VL brings: 🧠 Higher overall SixCat score. 📚 4× the native context capacity. 👁️ Powerful multimodal reasoning. 🎯 A compelling foundation for complex agent orchestration. These are measured deployment configurations with different inference engines, quantizations and speculative decoding strategies. This isn't an isolated architecture comparison. 𝗔𝗡𝗗 𝗧𝗛𝗔𝗧'𝗦 𝗘𝗫𝗔𝗖𝗧𝗟𝗬 𝗪𝗛𝗔𝗧 𝗠𝗔𝗞𝗘𝗦 𝗧𝗛𝗘𝗦𝗘 𝗥𝗘𝗦𝗨𝗟𝗧𝗦 𝗜𝗡𝗧𝗘𝗥𝗘𝗦𝗧𝗜𝗡𝗚! I care about what I can actually run and accomplish with these models on my own hardware! ━━━━━━━━━━━━━━━━━━━━ 🇺🇸 𝗣𝗔𝗥𝗧 𝟱: 𝗔𝗠𝗘𝗥𝗜𝗖𝗔𝗡 𝗔𝗜 𝗜𝗦 𝗣𝗨𝗦𝗛𝗜𝗡𝗚 𝗧𝗛𝗘 𝗢𝗡𝗘-𝗦𝗣𝗔𝗥𝗞 𝗙𝗥𝗢𝗡𝗧𝗜𝗘𝗥! One of the things I LOVE most about this release is seeing an American company like NVIDIA pushing open-weight local AI forward at this level! NVIDIA's published EXO Labs/local.ai results place Nemotron 3.5 Super VL on the intelligence-versus-task-completion-time Pareto frontier for a SINGLE 128GB DGX Spark. That's a published frontier evaluation, separate from my SixCat results. And then we have Bolt, with LiLiCorr DFlash delivering nearly 88 tokens per second in my single-stream test and over 195 output tokens per second on the balanced concurrent workload! Not on a rack of GPUs. 𝗢𝗡 𝗧𝗛𝗘 𝗦𝗣𝗔𝗥𝗞 𝗦𝗜𝗧𝗧𝗜𝗡𝗚 𝗜𝗡 𝗠𝗬 𝗛𝗢𝗠𝗘! This is what real progress looks like to me. Not simply releasing the BIGGEST model possible. But delivering: ✅ More useful intelligence per GB ✅ Faster task execution ✅ More concurrent work ✅ Larger native context windows ✅ Better deployment efficiency ✅ Open weights people can actually experiment with! 𝗧𝗛𝗘 𝗚𝗢𝗔𝗟 𝗜𝗦𝗡'𝗧 𝗝𝗨𝗦𝗧 𝗠𝗔𝗫𝗜𝗠𝗨𝗠 𝗜𝗡𝗧𝗘𝗟𝗟𝗜𝗚𝗘𝗡𝗖𝗘. 𝗜𝗧'𝗦 𝗠𝗔𝗫𝗜𝗠𝗨𝗠 𝗨𝗦𝗘𝗙𝗨𝗟𝗡𝗘𝗦𝗦! 🔥 ━━━━━━━━━━━━━━━━━━━━ ⚡ 𝗣𝗔𝗥𝗧 𝟲: 𝗧𝗪𝗢 𝗠𝗢𝗗𝗘𝗟𝗦. 𝗢𝗡𝗘 𝗦𝗣𝗔𝗥𝗞. Now you can probably see exactly where I'm going with all of this... 😏 What if we combine the strengths of BOTH models? 🧠 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 — 𝗠𝗔𝗦𝗧𝗘𝗥 𝗔𝗚𝗘𝗡𝗧 Handling large-context reasoning, multimodal understanding, complex decisions and orchestration. ⚡ 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗕𝗢𝗟𝗧 — 𝗦𝗨𝗕-𝗔𝗚𝗘𝗡𝗧 Handling fast delegated work, coding, tool execution and high-throughput tasks. The goal? 𝗧𝗪𝗢 𝗣𝗢𝗪𝗘𝗥𝗙𝗨𝗟 𝗠𝗢𝗗𝗘𝗟𝗦. 𝗢𝗡𝗘 𝗦𝗣𝗔𝗥𝗞. 𝗙𝗨𝗟𝗟𝗬 𝗟𝗢𝗖𝗔𝗟. Imagine the lead model working with a massive context window, coordinating specialized workers that can execute tasks quickly without every request needing the larger model! That is the kind of practical agentic system I want to build. Today's model-to-model speed and quality comparisons come from separate deployments. Simultaneous residency, shared resource limits and actual master-worker delegation are the next things to validate. And YES, I've already been experimenting with smaller EXL3 models to make this vision more practical! 👀 ━━━━━━━━━━━━━━━━━━━━ 🔬 𝗔 𝗤𝗨𝗜𝗖𝗞 𝗡𝗢𝗧𝗘 𝗢𝗡 𝗧𝗘𝗦𝗧𝗜𝗡𝗚 All tests were performed on my single home GB10 DGX Spark. I successfully tested Super VL's full native 1M-token context window on this hardware. Bolt's new LiLiCorr numbers are speed-only measurements using the same Bolt target and vLLM engine as the matching no-draft baseline. The speed results shown above are from the reported SixCat Speed workloads. Disk measurements are not peak runtime memory usage. My purpose is to share reproducible, real-world results rather than suggest either configuration represents its maximum possible performance. ━━━━━━━━━━━━━━━━━━━━ 💚 𝗔 𝗛𝗨𝗚𝗘 𝗧𝗛𝗔𝗡𝗞 𝗬𝗢𝗨 𝗧𝗢 @nvidiaai! Being given pre-release access to Nemotron 3.5 Super VL was an incredible opportunity, and I've genuinely enjoyed spending the past week pushing these models through their paces. NVIDIA isn't just releasing impressive models. They're helping make powerful AI more accessible to developers, researchers and enthusiasts who want to run these systems on hardware they can own. And that's something I'm incredibly passionate about. 𝗢𝗡𝗘 𝗦𝗣𝗔𝗥𝗞. 𝗠𝗢𝗥𝗘 𝗜𝗡𝗧𝗘𝗟𝗟𝗜𝗚𝗘𝗡𝗖𝗘. 𝗠𝗢𝗥𝗘 𝗦𝗣𝗘𝗘𝗗. 𝗠𝗢𝗥𝗘 𝗣𝗢𝗦𝗦𝗜𝗕𝗜𝗟𝗜𝗧𝗜𝗘𝗦. 🚀 🔗 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗦𝗨𝗣𝗘𝗥 𝗩𝗟 𝗡𝗩𝗙𝗣𝟰 huggingface.co/nvidia/NVIDIA… 🔗 𝗡𝗘𝗠𝗢𝗧𝗥𝗢𝗡 𝟯.𝟱 𝗕𝗢𝗟𝗧 𝟮𝟰𝗚𝗕 huggingface.co/nvidia/NVIDIA… 🔗 𝗡𝗩𝗜𝗗𝗜𝗔 𝗟𝗜𝗟𝗜𝗖𝗢𝗥𝗥 𝗗𝗙𝗟𝗔𝗦𝗛 𝗗𝗥𝗔𝗙𝗧𝗘𝗥 huggingface.co/nvidia/NVIDIA… 📖 𝗡𝗩𝗜𝗗𝗜𝗔 𝗧𝗘𝗖𝗛𝗡𝗜𝗖𝗔𝗟 𝗕𝗟𝗢𝗚 developer.nvidia.com/blog/nv… 𝗧𝗪𝗢 𝗠𝗢𝗗𝗘𝗟𝗦. 𝗢𝗡𝗘 𝗦𝗣𝗔𝗥𝗞. 𝗧𝗛𝗘 𝗘𝗫𝗣𝗘𝗥𝗜𝗠𝗘𝗡𝗧𝗦 𝗔𝗥𝗘 𝗝𝗨𝗦𝗧 𝗚𝗘𝗧𝗧𝗜𝗡𝗚 𝗦𝗧𝗔𝗥𝗧𝗘𝗗! 💚🔥