NVIDIA released Nemotron 3.5 Bolt 74B-A9B on October 15, 2026, giving developers, creators and small teams a downloadable model for local coding and visual assistants. Its 24GB target is not a whole-computer memory requirement: NVIDIA’s technical report sets out a text-serving estimate, with vision listed separately.
According to NVIDIA’s model card, Bolt accepts text, images and video and produces text, including requests to use software tools. It still needs an application to turn those requests into actions. This is the released 74B-A9B model, not the separate, smaller Bolt E4B that NVIDIA schedules for October 23.
What the 24GB target actually counts
NVIDIA’s technical report, Table 2 and section 8.2, estimates 23.83GB for text-only serving. That includes a reported 22.87GB subtotal for model weights and runtime working memory, plus a 0.96GB helper model intended to accelerate generation. The estimate assumes one sequence, 131,072 tokens of context and an 8,192-token prompt-processing chunk.
Context is the model’s working space: input, conversation history, tool results and the reply share it. The model card allows up to 256K tokens, but that ceiling is not the configuration used for the report’s 24GB estimate.
The vision encoder and projector—the components that process visual input—add a separately listed 1.68GB. Calculation from NVIDIA’s figures: 23.83 + 1.68 = 25.51GB for the listed components with vision. That is arithmetic, not a measured peak or a recommended system minimum. It does not provide an allowance for the operating system and every other application.
Choose a documented configuration before buying
The GGUF deployment guide provides two hardware-labelled configurations using llama.cpp, software that runs the model locally:
NVIDIA recipe | Hardware listed by NVIDIA | Context capacity |
|---|---|---|
28GB | RTX Spark with 64GB, or a compatible discrete RTX GPU with at least 32GB | 65,536 tokens |
32GB | RTX Spark with 128GB, or a compatible discrete RTX GPU with at least 48GB | 262,144 tokens |
Both recipes include the vision component and helper model, and allow one concurrent sequence. Their 28GB and 32GB labels are not recommendations for total system memory. Use the matching recipe’s settings rather than combining the smaller memory figure with the larger context limit.
For the alternative vLLM runtime, the main model card specifies a particular nightly build or newer, plus additional kernels and parsers. Its summary lists several GPU generations, while its detailed integration section names Blackwell SM120/SM121. Compatibility outside that detailed path remains unresolved in this guide; do not assume that enough memory alone establishes support.
What the speed results can tell you
NVIDIA reports 83.99 tokens per second with the helper model in its coding-throughput benchmark. Its published measurements used llama.cpp on Windows 11 and a Surface Laptop Ultra with RTX Spark and 128GB of memory. They do not establish performance on a computer with 24GB of total memory, or on your particular GPU.
The accompanying AgentPerf Local result answers a different question from whether an assistant completes a task correctly. Artificial Analysis, which develops the benchmark, describes it as replaying recorded agent conversations to measure serving speed, not output quality. A fast replay is not proof that generated code works.
Local does not settle privacy or cost
The GGUF client example separates the model’s tool request from the application that executes it. The practical implication: local model processing does not make connected web search, browsing or other services offline. Check where those tools send information before describing the complete assistant as private or offline.
NVIDIA lists the model for commercial and non-commercial use. The OpenMDW 1.1 licence grants fee-free rights subject to its terms, including retaining the agreement and applicable origin notices when distributing model materials. Hardware, electricity, administration and any paid connected tools remain costs. These sources do not establish a local-versus-hosted break-even price.
A practical buying rule
Already have compatible hardware: evaluate the matching recipe on your actual workload. Check memory, response time and correctness before relying on it.
Buying mainly for Bolt: the product name does not justify a universal 24GB-system recommendation. That does not prove every 24GB setup fails; compatibility still depends on the configuration.
If your work is text-only, compare that requirement separately before paying for visual capability. RohitAI’s guide to Liquid AI’s local text model covers another workload choice, not a head-to-head performance result.
Reporting note: This AI-assisted guide draws on NVIDIA’s release materials, model cards and technical report, the benchmark developer’s documentation and the licence, checked on October 15, 2026. RohitAI did not run the model or independently measure memory use, speed or accuracy.