Nemotron 3.5 Bolt 74B-A9B

// Overview

Nemotron 3.5 Bolt 74B-A9B is the local member of NVIDIA's Nemotron 3.5 generation, released on 15 October 2026 alongside the data-centre Nemotron 3.5 Super VL. NVIDIA built it to run everyday agentic tasks — multistep reasoning, tool calling and web browsing — on GeForce RTX PCs, RTX PRO workstations, DGX Spark and DGX Station, and timed it for RTX Spark laptops arriving the following day. The design goal is stated as a budget: fit in 24GB of memory and decode at 80 or more tokens per second.

Bolt starts from Super VL rather than being trained from scratch. NVIDIA's Star Elastic structured compression cuts the 120B-parameter, 12B-active parent down to 74 billion total and 9 billion active parameters, with expert top-K fixed at 9 instead of 22. The flagship 24GB release then quantises the routed and shared expert weights to IQ2_XXS at about 2.06 bits per parameter, keeps the Mamba layers in NVFP4 and the embeddings, attention and output layers in Q8_0, and recovers accuracy with quantisation-aware distillation. It keeps Super VL's hybrid Mamba-2/Transformer LatentMoE backbone and C-RADIO v4-H vision encoder, so it reads images and MP4 video up to two minutes long as well as text, with a 256K-token context shared between prompt, visual inputs, history and tool results.

NVIDIA ships three checkpoints under OpenMDW-1.1: the packed 24GB model, a 24GB GGUF for llama.cpp-style runtimes, and an unquantised BF16 checkpoint intended mainly for customisation and post-training. A LiLiCorr draft model for speculative decoding is published separately; NVIDIA reports a mean acceptance length of 4.68 tokens with greedy decoding. NVIDIA names llama.cpp, Ollama, LM Studio, Unsloth and vLLM as supported runtimes. A smaller sibling, Bolt E4B 4GB, was announced in the same post for 23 October 2026.

On NVIDIA's published table the 24GB release beats Qwen3.6-35B-A3B, Gemma 4 26B-A4B and Nemotron 3.5 Lightning on most rows — GPQA Diamond 86.43, Terminal-Bench 2.1 50.98, BrowseComp 66.75 — and NVIDIA calls it the #1 open model in its class on PinchBench, where it scores 87.82 to Qwen3.6-35B-A3B's 88.07 in the table. The trade is speed: on NVIDIA's AIPerf SPEED-Bench coding run it decodes at 83.99 tokens per second against 96.63 for Qwen3.6-35B-A3B and 100.56 for Gemma 4 26B-A4B, though well ahead of the dense Qwen3.8 27B at 28.26.

// Benchmarks

This model's scores

1. PinchBench87.82%
2. GPQA Diamond86.43%
3. BrowseComp66.75%
4. AA-LCR64.06%
5. MCP-Atlas52.2%
6. Terminal-Bench 2.150.98%
7. WebArena (vision)50.86%
8. SciCode41.02%
9. Humanity's Last Exam (no tools)25.3%
10. τ³-bench (Banking)17.01%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• A 74B-parameter Mixture-of-Experts model packed into roughly 24GB, so it fits on a single consumer RTX GPU or DGX Spark
• Vision-language input — images and up to two minutes of video — in a local model, inherited from Nemotron 3.5 Super VL
• Ahead of Qwen3.6-35B-A3B, Gemma 4 26B-A4B and Nemotron 3.5 Lightning on GPQA Diamond, Terminal-Bench 2.1, BrowseComp and WebArena in NVIDIA's table
• Permissive OpenMDW-1.1 license with 24GB, GGUF and BF16 checkpoints plus a LiLiCorr speculative-decoding drafter
• 256K-token context window for long agent sessions on local hardware

// Best for

• Local computer-use and browsing agents on RTX laptops and desktops where data should not leave the machine
• Offline coding and terminal agents on DGX Spark or an RTX PRO workstation
• Running a capable multimodal model alongside other applications within a 24GB memory budget
• Fine-tuning and post-training from the BF16 checkpoint before re-quantising for local deployment

// How to access

// Nemotron (open-weight) — every version

The full lineage of the Nemotron (open-weight) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.