NVIDIA releases Nemotron 3.5 Super: 26 AA points and 630 tokens/s

NVIDIA has released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B. In independent measurements by Artificial Analysis, Super VL scores 26 on the Intelligence Index. That is double the score of Nemotron 3 Super (13), while the model uses roughly 60% fewer output tokens per task. Bolt, the first model in the new family, scored 21 and landed within two points of the previous flagship, Nemotron 3 Ultra.

US leader, with caveats

AA ranks Super VL as the strongest American open-weight model. It edges out Thinking Machines' Inkling Small (25.7) and clearly beats Nemotron 3 Ultra (23) and Meta's Muse Glimmer (17). The global picture is less rosy: the UAE's K2 Horizon 375B A23B scores 31, and China's MiMo-V2.6-Pro leads with 46. The US still trails in the open-weight segment; the gap has just narrowed a little.

The biggest surprise is speed. On DeepInfra's pre-release endpoint with NVFP4 weights, the model delivers about 630 tokens/s with 100,000-token prompts. That is faster than any model scoring 20 or higher on the index that AA has measured at this length. Inkling Small manages 124 tokens/s, Nemotron 3 Ultra 235, and Nemotron 3 Super 140. The first token arrives in 1.8 seconds, against 4.4 for Ultra and 8.5 for Inkling Small. The measurement was taken on one provider's pre-release infrastructure, so production numbers may differ.

Token usage also improves on the predecessor: 46,000 per task against 114,000. K2 Horizon uses 52,000 and MiMo-V2.6-Pro 64,000. But on the intelligence-versus-tokens curve, Super 3.5 is still far from the frontier.

Key specs of Super VL:

• 120B parameters in total, 12B active
• 1M-token context
• input: text and images; output: text
• released in NVFP4; the gap to BF16 on the index is no more than 0.1 points, and no single component shifted by more than 2.6 points

Bolt: almost Ultra at one-seventh the parameters

Bolt 74B-A9B scores 21.4. That is fourth among models under 100B parameters and first among American models in that class. It pulls 8.6 points ahead of Nemotron 3 Super (12.8) with 39% fewer parameters. The gap to Ultra is about 1.6 points, even though Bolt has seven times fewer parameters (74B against 550B) and six times fewer active ones (9B against 55B). Its context window is 256,000 tokens, and its input is multimodal.

The results profile is uneven. Bolt beats Ultra on AutomationBench-AA (15.5% against 3.0%) and AA-Briefcase (912 against 874 Elo), but trails on AA-LCR (70% against 79%) and Humanity's Last Exam (25% against 28%). In other words, it handles agentic and office scenarios better than long context and hard knowledge questions.

The price of that compactness is verbosity. Bolt uses 82,000 tokens per task: 1.8 times more than Super VL and 2.9 times more than Ultra (28,000), though 28% fewer than Nemotron 3 Super.

Where this is useful in practice

Super VL looks like a logical choice for long agentic chains and work with very large contexts, where latency and throughput matter. The NVFP4 model loses almost no quality, so you can deploy it right away on hardware that supports the format, without guessing how much quantization cost you. Bolt suits teams whose inference costs are bound by model size and whose tasks are closer to workflow automation than to academic benchmarks. But when choosing, count the cost per completed task, not the price per token: Bolt uses more tokens, and savings on parameters may partly be eaten by long outputs. Before migrating, run your own workloads and repeat the speed measurements on your production provider.