Nemotron 3.5 Bolt 74B-A9B
Released Oct 15, 2026 · knowledge cutoff unpublished
- Status
- Superseded
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 262K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Strengths & weaknesses
Strengths
- Native 2-bit quantisation squeezes it into 24 GB, so it runs on a single RTX card or a DGX Spark rather than a server
- Keeps Super VL's vision tower, so it reads screenshots, documents and short video locally with no cloud call
- Day-one support in llama.cpp, Ollama, LM Studio, Unsloth and vLLM, so it drops into existing local setups
Weaknesses
- No hosted API at all — you run it yourself or not at all
- Very verbose: it generated more than twice the median output tokens on Artificial Analysis' index, which slows local runs noticeably
- The BF16 release is meant for fine-tuning, not serving — the 24 GB build you would actually deploy gives up a few points across NVIDIA's own table
74B total / 9B active Mamba-2–Transformer LatentMoE compressed from Nemotron 3.5 Super VL with NVIDIA's Star Elastic method (expert top-K cut from 22 to 9, the parent's vision tower reattached); released 2026-10-15 on Hugging Face as BF16, the 24GB native-2-bit build and GGUF variants. Benchmarks are NVIDIA's model-card figures for the BF16 weights (GPQA Diamond 88.83, Terminal-Bench 2.1 53.93, HLE no tools 28.73); the deployable 24GB build scores 86.43 / 50.98 / 25.30. The card's 'GDPVal-AA-V2' column (31.13) is not on the Elo scale this site tracks, so gdpvalAA is null; SWE-bench is not reported. Context is 'up to 256K' (Artificial Analysis lists 262K). availability is self-host and pricing null: Artificial Analysis states there is no publicly available API and no OpenRouter listing was found. AA Intelligence Index 21 (reasoning variant) with 230M output tokens against a 100M median, but no cost or speed measurement. Training data was collected through 2026-02-24 and post-training data runs to May 2026; NVIDIA states no single knowledge cutoff. A 4 GB sibling, Bolt E4B, is due 2026-10-23. Compression from Super VL is lineage, not replacement, so predecessorId is null.
Benchmarks3 of 10 reported
GPQA Diamond
Terminal-Bench 2.1
Humanity's Last Exam
MMLU-Proretired
SWE-bench Verified
SWE-bench Pro
AIME (latest)retired
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
Not researched
Licence
OpenMDW-1.1 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor