Nemotron 3.5 Bolt 74B-A9B
Key facts · verified Oct 18, 2026
Nemotron 3.5 Bolt 74B-A9B is a NVIDIA model released on 15 Oct 2026. Pricing: Free — open weights, self-hosted (no first-party API pricing). Its context window is 256K tokens. It is not yet ranked on LMArena's text leaderboard. Source: huggingface.co.
NVIDIA’s Nemotron 3.5 Bolt 74B-A9B, released 15 October 2026: an open vision-language model with 74 billion total and 9 billion active parameters, compressed from Nemotron 3.5 Super VL. It takes images, video and text, returns text, and supports up to 256K tokens of context, with JSON output, tool calling and switchable reasoning. A 24GB version uses native 2-bit quantisation to fit in 24GB of memory and runs with llama.cpp, Ollama, LM Studio, Unsloth and vLLM. Weights are published under the OpenMDW License 1.1.
Free — open weights, self-hosted (no first-party API pricing)
256K tokens
Not rated
Not ranked
Key Features
Full Specifications
Free — open weights, self-hosted (no first-party API pricing) | |
|---|---|
| Identity | |
| Developer | NVIDIA |
| Released | Oct 2026 |
| Status | GA |
| Licence | OpenMDW License Agreement 1.1 |
| Self-hostable | Yes — 24GB version (2-bit) for local GPUs; BF16 on A100, H100 or Blackwell |
| Cost | |
| Input price | Self-hosted — no first-party API price |
| Output price | Self-hosted — no first-party API price |
| Capacity | |
| Context window | 256K tokens |
| Capability | |
| Vision in | Yes — images and video |
| Audio in | No |
| Function calling | Yes |
| Structured output | Yes — JSON output |
| Extended reasoning | Yes — configurable via chat template (enable_thinking) |
| Access | |
| API | Not offered first-party |
| Fine-tuning | Yes — open weights permit it |