NVIDIA released the Nemotron 3.5 Bolt 74B-A9B 24GB model on October 15, 2026, designed to run local agentic tasks on RTX/DGX Spark hardware, with a 256K context window and 84 tokens per second performance.
The model’s release coincided with the launch of NVIDIA RTX Spark laptops on October 16, 2026, as noted by Kie.ai.
Model Specifications and Performance
The Bolt 74B-A9B model, derived from the Nemotron 3.5 Super VL (120B-A12B), uses a hybrid Mamba-2/Transformer mixture-of-experts architecture. It maintains 9 billion active parameters per token while compressing the total to 74 billion through NVIDIA’s Star Elastic framework. This compression balanced size and capability by reducing the expert top-K. The model achieves 84 tokens per second on RTX Spark hardware, according to NVIDIA’s benchmarks.
This is interesting.Nemotron-3.5-Bolt-74B-A9B is not a seperate model trained from scratch. It's actually a compressed version of larger Nemotron-3.5-SuperVL-120B-A12BNVIDIA uses its Star Elastic framework to shrink the model from 120B – 74B, reducing active parameters from… https://t.co/1Mw1AcRHg9 pic.twitter.com/s8YBnN5ey1
— https://x.com/ItsmeAjayKV/status/2110833345835999678
Kie’s analysis highlights the model’s 256K context window and 2-bit quantization, which allows it to operate within 24GB of memory. It ranks #1 on PinchBench among models with similar memory footprints and scores 21 on the Artificial Analysis Intelligence Index. The model’s open weights are available under the OpenMDW License v1.1, enabling commercial use and customization via NVIDIA NeMo.
Comparison with Other Models
When compared to other open models, the Bolt 74B-A9B outperforms Qwen3.6-35B-A3B on BrowseComp by 18.1 points but trails newer Qwen 3.8 models on independent benchmarks. The model’s Terminal-Bench 2.1 score of 50.98 further illustrates its capabilities. NVIDIA’s technical blog emphasizes that the Bolt model is designed for local agent workloads, contrasting with the Super VL’s enterprise focus and the Lightning model’s high-throughput tasks.
Nvidia has delivered a Qwen3.6-35B killer and it's a 74B → 24GB model, actually, today. And now we have real numbers. 👀Nemotron 3.5 Bolt 74B-A9B🧠 74B total / 9B active💾 22.1GB quantized weights🔢 ~2.06-bit experts + QAT📚 up to 256K context👁️ text + images + video🤖… pic.twitter.com/HoVUCax36n
— https://x.com/TeksEdge/status/2110792828041167202
Kie’s comparison table shows the Bolt 74B-A9B’s parameters (74B total, 9B active) versus the Super VL (120B total, 12B active). The Bolt model’s 2-bit quantization allows it to run on consumer-grade hardware like GeForce RTX and Jetson devices, as noted in the Kie article.
Release Timeline and Availability
The Nemotron 3.5 Bolt 74B-A9B 24GB became available on October 15, 2026, with the smaller Bolt E4B 4GB version scheduled for October 23, 2026. NVIDIA’s blog states that the Super VL and Bolt 74B-A9B are now available, while the E4B version will follow. The Bolt E4B is designed for conversational and physical agents on GeForce RTX, Jetson, and RTX/DGX Spark devices, as outlined in Kie’s report.
NVIDIA’s NIM packages provide production-ready microservices with enterprise-grade support, updated every two weeks. The model runs on NVIDIA’s full hardware stack, from DGX Spark to RTX PRO systems, as detailed in the NVIDIA blog.
Details E4B release to expand edge AI
The upcoming Bolt E4B 4GB release on October 23, 2026, will expand NVIDIA’s edge AI offerings.
Run Nemotron 3.5 Bolt on your DGX Spark ⚡️This is a very interesting model. It could replace the legendary Qwen3.6 35B for me, but I haven't tested it much yet.- 256k context, 16 concurrent streams- Uses ~52 GB of system memory- 71 tok/s on prose, single steam- 166 tok/s…
— https://x.com/MiaAI_lab/status/2111103752824385843
More on Nvidia
NVIDIA just dropped Nemotron 3.5 Bolt, a 24GB compressed version of their new Super VL (74B total, 9B active, native 2-bit). Their own launch chart puts Qwen3.8 27B above it on intelligence and Bolt way ahead on speed.So I ran both on a 3090 and a 5090. 480 trials, same seeds,… https://t.co/qSnZ3gsutX pic.twitter.com/LqPYQKc8lu
— https://x.com/0xdippo/status/2111090423670215042
Nemotron 3.5 Bolt, 74B with 9B active, in 22 Gb, on one DGX Spark with vLLM.29 tok/s with no drafter. NVIDIA's LiLiCorr drafter takes it to 85, but then repeated prompts reuse no cache, so I run it without.Tool calling 23 of 24. My 60-item agent suite 54 of 60. Reliability 13… pic.twitter.com/5flnp2ihfX
— https://x.com/ScottLeimroth/status/2111220857179971987
And this one fits into 24GB of memory.Introducing Nemotron 3.5 Bolt 24GB ⚡️ https://t.co/yPsBt9N58A pic.twitter.com/SOiBgeMMjD
— https://x.com/NVIDIAAI/status/2110824076218605922
NVIDIA Nemotron 3.5 Bolt 24GB puts powerful local agents within reach. ⚡74B-A9B parameters packed into 24GB of memory, balancing intelligence and speed while running alongside your apps. Optimized for RTX Spark and ready for your next project. pic.twitter.com/wzbu8aN8ln
— https://x.com/NVIDIARTXSpark/status/2110813633697694003
And another one, this time for your laptop. ⚡️Alongside Nemotron 3.5 Super VL, meet NVIDIA Nemotron 3.5 Bolt. Built for local AI agents on RTX Spark with 74B parameters in just 24GB of memory.Check it out ⬇ https://t.co/H8NQF2eEwQ
— https://x.com/nvidia/status/2110804342487810232