Nvidia bills Nemotron Bolt #1 on PinchBench, its card scores it #2

Nvidia’s launch post for Nemotron 3.5 Bolt 74B calls it “the #1 open model in its class on PinchBench, and the #2 open model for BrowseComp.” The only table Nvidia published comparing Bolt with other models says the reverse. On the model’s own Hugging Face card, Bolt scores 87.82 on PinchBench against 88.07 for Qwen3.6-35B-A3B, and 66.75 on BrowseComp against 48.74. It’s second on one and first on the other, not the pairing the blog describes.

Nvidia released three open models on 15 October 2026. Nemotron 3.5 Super VL is a 120B mixture-of-experts model that activates 12B parameters per token, with a 1M-token context and an OpenMDW-1.1 licence. Bolt 74B-A9B 24GB is a 2-bit quantized sibling built to sit in 24GB of memory, and a third model, Bolt E4B 4GB, is due 23 October. That matters because the card is the only place Nvidia shows Bolt next to a rival.

The result is a table that mostly flatters Bolt. It compares the model against three rivals across 13 benchmarks. Bolt leads on 10 of them and trails on three, which is a strong showing on its face. The two Nvidia chose to put in the blog are the ones where the ordering got flipped.

Benchmark	Nemotron 3.5 Bolt 74B-A9B (IQ2)	Qwen3.6-35B-A3B BF16	Gemma4 26B-A4B BF16
PinchBench	87.82	88.07	74.70
BrowseComp	66.75	48.74	26.30
MCP-Atlas	52.20	51.20	not reported
AA-Omniscience accuracy	16.90	19.47	22.17

The catch is that the MCP-Atlas row carries its own asterisk. Nvidia’s footnote reads that Qwen’s reported score is 62.8, and that Nvidia measured 51.2 on the 500 public tasks. Which means Bolt’s 52.20 clears Qwen only on Nvidia’s own re-measurement. At the number Qwen published, Bolt loses that row by more than 10 points.

The reported MCP-Atlas score for Qwen3.6-35B-A3B-BF16 is 62.8; we measured 51.2 on the 500 public MCP-Atlas tasks.

Nvidia, Nemotron 3.5 Bolt 74B-A9B 24GB model card

What the speed numbers are measured on

The same pattern shows up on the speed figures. Super VL’s headline claim is two different claims wearing one number. Nvidia’s blog says it’s “5x faster than the next best.” The developer post attributes the 5x to throughput on a single GPU against open models in the same size category, then gives a separate figure of 4x for task completion. SiliconANGLE reported it as outputs produced “nearly five times faster.”

The serving claim narrows the same way once you read the conditions. The blog says Dynamo lets organisations “serve up to 8x more users with 85% lower latency.” The developer post states the condition: 16 B200 GPUs, at 50 tokens per second per user, with KV-aware routing and cache offload. The 85% is first-token latency specifically, not latency across the board. That’s a 16-GPU data centre configuration, not a desktop one, which is the opposite end of the range from the RTX Spark hardware the Bolt models target.

Even so, quantization costs something, and the card shows where. Nvidia ships Super VL in NVFP4 and says quantization-aware distillation closes the accuracy gap. On 33 of the 42 rows on the Super VL card, the two versions land within two points of each other. The catch is the offline WebVoyager run, where they don’t: BF16 scores 77.72 and NVFP4 scores 66.31, the widest gap on the card.

Customer results are forecasts, not measurements

The result is the same on the customer side. Nvidia named a long list of companies post-training Super VL, and the developer post puts numbers on three. Scale AI built a vendor credit agent for Skydance where “inference cost for this workload will decrease by 40x and latency by 5x.” Ramp “expects up to 50% higher throughput and 10x lower latency” on receipt extraction. AI21 and Alice improved accuracy by 41% on threat analysis.

Which means the ordering of tenses does the work here. The AI-generated summary at the top of that same page compresses all three into one line: those organisations are “achieving improvements such as 40x lower inference cost, 10x lower latency, and 40% higher accuracy.”

Two of the three figures are forecasts in the body text. The 10x belongs to Ramp’s expectation, not to a measured result, and the 41% has become 40%. Nvidia’s own disclaimer under the summary says AI-generated content may summarise incompletely.

Even so, none of this makes Bolt a weak model. A 74B model squeezed to 24GB, with its experts at 2.0625 bits per parameter, beating BF16 rivals on 10 of 13 rows, is a real engineering result, and it runs through llama.cpp, Ollama and vLLM like any other local checkpoint. The gap is between the card and the press copy, which is the same gap we found when Reflection AI’s Beam trailed two rivals on benchmarks it chose itself.

Bolt E4B 4GB lands on 23 October, and its card is the thing to read first. Even so, Nvidia publishes its comparison tables, which is more than most labs do. It just doesn’t always read them before writing the blog.