Nemotron 3.5 Super VL and Bolt 74B-A9B added to BenchGecko · NVIDIA open multimodal MoE models
NVIDIA released Nemotron 3.5 Super VL (120B total, 12B active, up to 1M context) and Nemotron 3.5 Bolt 74B-A9B (24GB, 256K context) as open weights on 2026-10-15.
What landed
Nemotron 3.5 Super VL is an open vision-language model from NVIDIA, released on 2026-10-15 on Hugging Face, OpenRouter, ModelScope and as NVIDIA NIM. NVIDIA released Nemotron 3.5 Bolt 74B-A9B the same day, a compressed version built to run locally in 24GB of memory. A third model, Nemotron 3.5 Bolt E4B 4GB, is scheduled for 2026-10-23 and is not listed yet.
Super VL specs
From NVIDIA's model card:
- Parameters · 120B total, 12B active per token
- Architecture · Mamba-2 and Transformer hybrid MoE with 512 routed experts per MoE layer, a C-RADIO vision encoder and a multi-token prediction head
- Context · up to 1M tokens, shared by prompt, visual inputs, tool results and output
- Inputs · text, images and video (mp4 up to 2 minutes)
- Output · text
- Reasoning · switched on or off through the chat template
- Licence · OpenMDW-1.1, open weights plus open datasets and training recipes
NVIDIA says the model was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples. BenchGecko stores the context field as 1,000,000 tokens.
Super VL vendor-reported benchmarks
From the BF16 model card. Vendor-reported, not BenchGecko tests.
- GPQA Diamond · 89.52
- LiveCodeBench v6 · 88.22
- SWE-bench Verified · 71.10
- SWE-bench Pro · 51.78
- Terminal-Bench 2.1 · 66.70
- BrowseComp · 78
- OSWorld-Verified · 51.14
- MMLongBench-Doc · 62.82
- CharXiv reasoning · 80.82
- HLE, text-only · 31.79
NVIDIA also cites up to 5x higher throughput than other models in its class, based on the Artificial Analysis Intelligence Index leaderboard.
Bolt 74B-A9B specs
- Parameters · 74B total, 9B active, compressed from Super VL
- Precision · native 2-bit expert weights with quantization-aware training, so the packed model fits in 24GB
- Context · up to 256K tokens, stored as 262,144 on BenchGecko
- Inputs · text, images and video
- Output · text
- Licence · OpenMDW-1.1
Vendor-reported results from the Bolt model card:
- GPQA Diamond · 86.43
- Terminal-Bench 2.1 · 50.98
- BrowseComp · 66.75
- PinchBench · 87.82
- HLE, no tools · 25.30
- SciCode · 41.02
The same card lists Nemotron 3.5 Lightning at 75.57 on GPQA Diamond and 23.46 on Terminal-Bench 2.1.
Pricing on BenchGecko
NVIDIA does not publish a per-token price for either model. OpenRouter lists a free Super VL endpoint. BenchGecko leaves the pricing fields empty until a paid provider price is listed. For reference, Nemotron 3 Super, the earlier text-only model of the same size, is listed at $0.08 input and $0.45 output per M tokens.
Sources
- NVIDIA announcement
- Super VL model card
- Bolt 74B-A9B model card
- BenchGecko model pages: /model/nemotron-3-5-super-vl and /model/nemotron-3-5-bolt-74b-a9b
Source: BenchGecko
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.