Nvidia Releases Nemotron 3.5 Bolt, a 74B Vision Model Packed Into 24GB at 2 Bits
A.I. / news
Nvidia Releases Nemotron 3.5 Bolt, a 74B Vision Model Packed Into 24GB at 2 Bits
The Hugging Face card gives a 256K context and the OpenMDW-1.1 licence; every benchmark score is Nvidia's own.
Nvidia published Nemotron 3.5 Bolt 74B-A9B-24GB on Hugging Face on Oct. 15, a 74-billion-parameter vision-language model squeezed into a 24GB checkpoint with native 2-bit weights. Its model card lists a licence, OpenMDW-1.1, that it says allows commercial and non-commercial use.
The model is compressed from Nvidia's larger Nemotron 3.5 Super VL 120B-A12B, according to the Hugging Face model card. Nvidia describes it as built for everyday agentic tasks run locally.
What the card specifies
Bolt is a mixture-of-experts model with 74 billion total parameters and 9 billion active per token. The card describes a hybrid of Mamba-2 and Transformer layers with a LatentMoE design and a C-RADIO v4-H vision encoder. It takes text, images and video and returns text, with a context of up to 256,000 tokens.
The "24GB" in the name is the packed checkpoint size. Nvidia says it uses native 2-bit quantisation, trained in rather than applied afterwards, so the model fits in about 24GB of memory and can run beside other applications. Seventy-four billion parameters at 2 bits is roughly 18.5GB before the cache and vision encoder, which is consistent with that figure.
| Spec | Bolt 74B-A9B | Super VL 120B-A12B |
|---|---|---|
| Total / active parameters | 74B / 9B | 120B / 12B |
| Context | 256K tokens | 1M tokens |
| Weight precision | 2-bit (IQ2) | NVFP4 build on Modal |
| Licence | OpenMDW-1.1 | OpenMDW-1.1 |
The parent column draws on the Modal library page for the NVFP4 build of Super VL, which lists the same parameter counts, a one million token context and the same licence.
Benchmarks are Nvidia's own
The card reports scores for the 2-bit model on five tests. All are self-reported by Nvidia.
| Benchmark | Bolt (IQ2) |
|---|---|
| GPQA Diamond | 86.43 |
| BrowseComp | 66.75 |
| Terminal-Bench 2.1 | 50.98 |
| WebArena (vision) | 50.86 |
| Humanity's Last Exam, no tools | 25.30 |
Nvidia lists Qwen3.6-35B-A3B, Gemma4 26B-A4B and Nemotron 3.5 Lightning 30B-A3B, all in BF16, as comparison points. The card shows Bolt ahead on most rows but not all. Gemma4 leads on AA-Omniscience accuracy.
No independent evaluator is named. The comparison also sets a 2-bit model against 16-bit rivals, so a score gap says little about how much quality the quantisation cost against Bolt's own parent.
Hardware and what does not add up
The card lists deployment on vLLM with Nvidia Ampere (A100), Hopper (H100 80GB) or Blackwell GPUs. Tested hardware includes H100, H200, GB200, RTX 5090, DGX Spark and RTX Spark. It names no minimum for consumer cards.
One field on the page disagrees with the text. The card's metadata shows a model size of 22 billion parameters, while the text says 74 billion total. The card does not address the gap. The card gives training data dates of December 2025 to May 2026.
For comparison, the site has covered Liquid AI's LFM2.5-24B-A2B, a small mixture-of-experts release, and the far larger Meta Muse Spark Open.
Nvidia has not said whether it will publish a technical report for Bolt.
Sources
More in A.I.
- 01Vals Test: Three of Six AI Agents Threatened Court Over a £350 Train RefundTold to try everything to recover a refund the rules no longer allowed, Grok 4.7 sent 187 messages in 15 simulated days while Opus 5.5 sent eight in twelve weeks.
- 02MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark TablesOpenBMB's first mixture-of-experts vision model claims a win over Qwen3.6 35B-A3B in prose and charts, and 13 tokens a second on an iPhone, but the model card prints no scores.
- 03Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own TableGoogle's October 13 launch post leads with a Vals Index ranking; the comparison table behind it shows Argon behind GPT-6 Astra or Opus 5.5 on both Terminal-Bench tests and on OSWorld 2.0.
- 04StarSkirmish Has Nine Models Write StarCraft Bots in One Hour, and the Best Wins 7% Against Top Human BotsGemini 4 Argon lands in a three-way tie at 37 to 39, but it ran at a lower reasoning setting than most rivals and its runs vary the most.