Nemotron 3.5 Bolt 74B-A9B
NVIDIA·
LLMsopen-weightWorkstation
Overview
A 74B / 9B-active hybrid MoE built for local agents, shipped by NVIDIA as a 2-bit (IQ2) GGUF trained with quantization awareness so the whole checkpoint fits in 24 GB — the RTX Spark and DGX Spark target. Takes text and image input at a 262K context; runs through llama.cpp, Ollama, LM Studio, Unsloth and vLLM. NVIDIA describes it as fully open (weights, data, recipes) but the license name was not confirmable at intake. Note the hardware class shown here is computed at Q4 like every other entry; the 24 GB figure depends on NVIDIA's 2-bit build.
Capabilities and innovations
74B total / 9B active parameters (MoE)2-bit quantization-aware build fits in 24 GBText and image input262K token context windowReasoning modelllama.cpp, Ollama, LM Studio, vLLM supportQuantization-aware 2-bit release as the primary checkpoint
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 86.43% | vendor GPQA Diamond 86.43, measured on the 2-bit IQ2 build (comparison models in BF16). Card not fetchable from this environment (huggingface.co egress-blocked); value taken from consistent secondary transcriptions of the card. |
| Humanity’s Last Exam (no tools) | 24.65% | third party HLE 24.65% no-tools, Artificial Analysis Intelligence Index component (accessed 2026-10-15; Intelligence Index 21.4). NVIDIA's card reports 25.30. |
| Terminal-Bench 2.x | 50.98% | vendor Terminal-Bench 2.1 50.98 on the 2-bit IQ2 build. Card not fetchable from this environment (huggingface.co egress-blocked); value taken from consistent secondary transcriptions of the card. |
- SWE-bench Pro: not reported by the vendor — The card reports PinchBench, MCP-Atlas, BrowseComp and WebArena instead of SWE-bench.
- MMLU-Pro: not reported by the vendor
- Terminal-Bench 4.0: not reported by the vendor — Artificial Analysis scores 0 on its own Terminal-Bench 4.0 run, a harness on which most frontier models also score 0; not comparable with the vendor Harbor values in the field.
Architecture and hardware
- Parameters
- 74B total, 9B active per token (MoE)
- Estimated VRAM at Q4
- ~43 GB, Workstation class
- Quantization formats
- GGUF (IQ2)
- Recommended runtime
- Ollama
Links
More from NVIDIA
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.