August 11, 2026
Benchmarks, beef, and bot bragging
Nvidia Nemotron 3.5 Lightning
Nvidia drops a speedy new AI, and the comments instantly turn into a benchmark cage match
TLDR: Nvidia released Nemotron 3.5 Lightning, a fast new open AI model for commercial use with unusually open training details. Commenters immediately split into two camps: one cheering the openness and fresh design, the other asking why rival Qwen still seems to win the numbers game.
Nvidia has unveiled Nemotron 3.5 Lightning, a new openly released AI model meant for business use, chatbots, coding help, and so-called AI agents. On paper, it’s a big flex: long memory, faster text generation, and a training recipe Nvidia says is open, reproducible, and ready for developers to use. But in the comments, the real show begins: this wasn’t just a launch, it was a full-on scoreboard brawl.
One crowd was impressed that Nvidia is pushing a newer design instead of the usual style of AI model. Commenters called the use of Mamba 2 “really interesting,” basically treating it like the upstart challenger finally landing punches on the old champion. Another camp, though, came in swinging with the classic internet response: cool story, but how does it compare to Qwen? One user bluntly said it was “crazy” how well Qwen’s rival model is still holding up, turning the thread into a nerdy version of sports fans arguing stats after a playoff game.
Then came the nuance squad. A few commenters argued Nemotron may look behind on some score charts, but could be less ‘benchmaxxed’ and stubborn in real use — which is a very polite way of saying, “Maybe your spreadsheet doesn’t tell the whole story.” Others praised Nvidia for something rarer than top scores: publishing not just the model, but the data and recipe too. The vibe? Respect for the openness, mild panic about the rankings, and a lot of side-eye at benchmark bragging. Even the shortest comment — “Nice cadence of releases” — read like a calm voice whispering while everyone else flipped tables.
Key Points
- •NVIDIA released NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, an open-weight LLM available for commercial use under the OpenMDW-1.1 license.
- •The model uses a hybrid Mixture-of-Experts architecture with interleaved Mamba-2 and MoE layers plus select attention layers, with 30B total parameters and 3B active parameters.
- •NVIDIA states the model was pre-trained on more than 20T tokens, supports up to 1M context length, and has pre-training and post-training data cutoffs of September 2025 and May 2026 respectively.
- •The article includes a deployment recipe for DGX Spark using vLLM, DSpark speculative decoding, Marlin, FlashInfer, FP8 KV cache, and tool/reasoning parser settings.
- •NVIDIA published benchmark results and reproducibility materials through NeMo Gym and NeMo Evaluator SDK, including scores such as 81.62 on MMLU Pro, 75.57 on GPQA Diamond, and 52.80 on SWE-bench Verified for the NVFP4 checkpoint.