Released in October 2026, Bolt reaches 83.99 tokens per second on an RTX Spark laptop in NVIDIA's llama.cpp test. It suits people running coding, browsing and research agents locally, and it decodes slower than the Qwen and Gemma rivals NVIDIA tested alongside it.
What changed
Nemotron 3.5 Bolt 74B-A9B 24GB is NVIDIA's open vision-language model that scores 86.43 on GPQA Diamond in NVIDIA's own testing while fitting into 24GB of memory. It activates 9B of its 74B parameters per token, stores experts in 2-bit form, and takes text, image and video input.
Where it sits
- 86.4%% correctGPQA DiamondHigher is better#33 / 59peer median 87.4%per source, see benchmark scores
Mid-pack on GPQA Diamond (rank 33 of 59), and one of 82 whose vendor states it does not train on customer data.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: NVIDIA · Family: Nemotron 3.5
Context window: 262,144 tokens
Input modalities: text, image, video, tool-calls · Output: text, tool-calls
About Nemotron 3.5 Bolt 74B-A9B 24GB
Nemotron 3.5 Bolt 74B-A9B 24GB is an open-source vision-language model that NVIDIA published on Hugging Face on October 15, 2026, alongside the larger Nemotron 3.5 Super VL. NVIDIA says a smaller Bolt E4B follows on October 23. Bolt has 74 billion total parameters and 9 billion active per token. Its backbone mixes Mamba-2 layers, latent mixture-of-experts layers and attention, and a C-RADIO v4-H encoder reads images and video frames. NVIDIA compressed it from the 120B-A12B Super VL parent with a router-guided search it calls Star Elastic, cutting the expert top-K from 22 to 9, then quantized every routed and shared expert to IQ2_XXS at 2.0625 bits per element so the checkpoint packs into 24GB. The older Nemotron 3 Nano 30B A3B and the text-only Nemotron 3.5 Lightning are the nearest siblings in the same family, and HokAI's NVIDIA model directory lists the rest.
NVIDIA's model card compares the 2-bit checkpoint with Qwen3.6-35B-A3B, Gemma 4 26B-A4B and Nemotron 3.5 Lightning, all three run in BF16. Bolt posts 86.43 on GPQA Diamond against 83.40 for Qwen3.6 and 79.61 for Gemma, 50.98 on Terminal-Bench 2.1 against 44.38 and 37.22, and 66.75 on BrowseComp against 48.74 and 26.30. It also reports 25.30 on Humanity's Last Exam with tools off, 52.20 on MCP-Atlas and 50.86 on the vision WebArena test. It does not win every row. Qwen3.6 leads PinchBench, 88.07 to 87.82, Gemma 4 leads AA-Omniscience accuracy, 22.17 to 16.90, and Lightning has the better AA-Omniscience non-hallucination score, 68.81 to 61.27. NVIDIA's blog calls Bolt the number one open model in its class on PinchBench, which holds only inside the similar-memory-footprint group NVIDIA chose, because its own table puts a BF16 Qwen3.6 slightly ahead. NVIDIA ran every one of these evaluations itself, and it re-measured Qwen3.6 on MCP-Atlas at 51.2 instead of the 62.8 Qwen reported. Artificial Analysis independently gives the model an Intelligence Index of 21 (v4.3.2) and records 230M output tokens used to run the index, but its page links the BF16 weights and lists no public API, so it is unclear whether the score describes the 2-bit checkpoint.
The context window reaches 256K tokens, and NVIDIA's own vLLM recipe sets 262,144. The prompt, images, conversation history, tool results and reply share that window, and NVIDIA notes that a deployment recipe may use less to fit GPU memory. Inputs are text, images and video, and output is text. Video means mp4 clips up to 2 minutes, sampled at up to 1 frame per second (128 frames) for 1080p or 2 frames per second (256 frames) for lower resolutions. Artificial Analysis lists only text and image input, and NVIDIA's vision adapter stage used images only. Reasoning is switchable in the chat template with enable_thinking on or off, plus a medium effort option, and the model supports JSON output and tool calling. English and code are the primary languages, with Spanish, French, German, Italian and Japanese also supported.
Speed was the design target: NVIDIA aimed for at least 80 tokens per second and reports 83.99 on AIPerf SPEED-Bench-Coding. That run used llama.cpp on an RTX Spark laptop (Surface Laptop Ultra, 128GB, Windows 11 ARM) with the 1.13B-parameter LiLiCorr drafter proposing seven tokens at a time. In the same table Qwen3.6 35B-A3B at Q4_K_M reached 96.63, Gemma 4 26B-A4B reached 100.56 and the dense Qwen3.8-27B reached 28.26. On AA AgentPerf Local, which replays agent trajectories, Bolt took 600.80 seconds against 453.97 for Qwen3.6, 480.68 for Gemma and 1043.28 for Qwen3.8 27B, so it finishes roughly a quarter to a third slower than the two faster rivals despite scoring higher on most accuracy rows of NVIDIA's table. HokAI's record for Qwen3.8-27B lists 89.2 on GPQA Diamond and 73 on Terminal-Bench 2.1, above Bolt on both, so that dense model wins on raw scores while Bolt wins on speed. NVIDIA lists support in llama.cpp, Ollama, LM Studio, Unsloth and vLLM, while Hugging Face hosts the files. The vLLM recipe runs one sequence at a time on one GPU, and NVIDIA tested on H100, H200, GB200, RTX 5090, DGX Spark and RTX Spark.
The weights are open under OpenMDW-1.1, and NVIDIA states the model is ready for commercial or non-commercial use. There is no per-token price because there is no hosted endpoint to bill: Artificial Analysis shows no public API, and OpenRouter's model list did not include Bolt when HokAI checked on October 20, 2026. Your cost is the machine. NVIDIA's compression search targeted a footprint under 24GB at batch size 1 and 128K context, and its blog names RTX Spark laptops (arriving October 16), GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations. Bigger servers that want the parent model's behavior can use Super VL, which NVIDIA lists on Hugging Face, OpenRouter, ModelScope and NIM.
NVIDIA's card gives two different token counts for the compression stage, 120 billion in the training methodology and 80 billion in the dataset section, so treat the figure as unsettled. Post-training data runs to May 2026. The card says Kimi K3 generated synthetic data that targets benchmarks such as GPQA Diamond and Terminal-Bench 2.1, so some headline scores are also training targets, and a small share of trajectories from GLM 5.2 targeted MCP-Atlas. HokAI has pages for both teachers: GLM-5.2 and Kimi K3. Reinforcement learning used the Nemotron 3 Ultra reward model and agentic environments for MCP and software tasks. Training data was filtered with a Gemma-3 4B-based guard model, image datasets were scanned for CSAM, and NVIDIA's privacy subcard says no personal data and no user prompts were used for training. The card publishes no refusal-rate benchmark and names no outside red-team partner.
Bolt fits people who want a coding, research or browsing agent running on their own machine with no data leaving it, on hardware with roughly 24GB of memory to spare. Skip it if you need the fastest decode (Qwen3.6 and Gemma 4 beat it in NVIDIA's test), strong factual recall (its AA-Omniscience accuracy is 16.90), a hosted API with an SLA, or multi-user serving. Every benchmark above is vendor-run on a 2-bit checkpoint that outsiders had not reproduced when this page was written.
Pricing
Nothing is billed per token: the files cost nothing to fetch and no hosted endpoint turned up at upload. The real cost is the 24GB-class machine that runs it. A sibling with listed API prices is [Nemotron 3.5 Lightning](/hub/models/nemotron-3.5-lightning-30b-a3b).
| Tier | Rate |
|---|---|
| Self-hosted (24GB checkpoint) | Free weights under OpenMDW-1.1; cost is your own hardware |
| Hosted API | No hosted endpoint found on 2026-10-20; Artificial Analysis lists no public API |
Key Features
- Star Elastic compression: NVIDIA shrank the 120B-A12B Super VL parent to 74B-A9B by choosing expert counts per layer and cutting expert top-K from 22 to 9.
- 2-bit expert weights: All routed and shared experts use IQ2_XXS at 2.0625 bits per element, with quantization-aware distillation to recover accuracy.
- LiLiCorr speculative drafter: A 1.13B-parameter drafter proposes seven tokens per step and reached a mean acceptance length of 4.68 with greedy decoding at concurrency 64 in vLLM.
- Text, image and video input: A C-RADIO v4-H encoder feeds images and mp4 clips up to 2 minutes into a long-context window.
- Switchable reasoning: The chat template turns thinking on or off with enable_thinking and offers a medium effort setting.
Pros
- Leads NVIDIA's browsing rows against both the Qwen and Gemma rivals it was tested with, which matters for research agents.
- Keeps image and video input inside a 24GB package, so one local model covers text and visual agent steps.
- Ships weights, GGUF builds and a speculative drafter under one open license, so there is no lock-in on the model.
Cons
- Slower than both the Qwen and Gemma rivals on decode and on agent replay in NVIDIA's own test.
- Weaker factual recall, with AA-Omniscience accuracy below both rivals on the model card.
- Benchmarks come from NVIDIA alone, and the card contradicts itself on compression training tokens.
Benchmarks
- MCP Atlas: 52.2% vendor-reported · 20 Oct 2026 — Tasks solved by calling external tools over the Model Context Protocol, % solved.
- GDPval-AA v2: 28.47 vendor-reported · 20 Oct 2026 — Real knowledge-work deliverables judged against professionals, run by Artificial Analysis.
- GPQA Diamond: 86.4% vendor-reported · 20 Oct 2026 — PhD-level science questions that are hard to search for, % correct.
- Terminal-Bench 2.1: 51.0% vendor-reported · 20 Oct 2026 — Multi-step tasks completed in a real command line, % solved.
- Humanity's Last Exam: 25.3% vendor-reported · 20 Oct 2026 — Expert-written questions across many fields, % correct.
- AA Intelligence Index: 21 cited: Artificial Analysis · 20 Oct 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What does it cost to run Nemotron 3.5 Bolt?
Nothing per token: the weights are free under OpenMDW-1.1, and no vendor-hosted endpoint was found on October 20, 2026. Your cost is a machine with about 24GB of memory to spare, such as an RTX Spark laptop. For scale, OpenRouter lists the hosted Nemotron 3.5 Lightning at $0.07 input and $0.20 output per 1M tokens, and Nemotron 3.5 Super VL at $0.15 and $0.50.
Is Nemotron 3.5 Bolt better than Qwen3.6 35B-A3B?
It depends on the job. Bolt leads Qwen3.6 on browsing (BrowseComp 66.75 to 48.74) and terminal work, but Qwen3.6 edges it on PinchBench (88.07 to 87.82) and decodes faster, at 96.63 tokens per second in NVIDIA's speed test. The accuracy table runs Qwen3.6 at full precision while the speed table uses a 4-bit quant, and no outside group has reproduced either yet.
Is Nemotron 3.5 Bolt open source?
NVIDIA releases it as open weights under the OpenMDW-1.1 license and describes it as ready for commercial or non-commercial use. The Hugging Face repository carries the 24GB checkpoint, with GGUF, BF16 and drafter variants alongside, and NVIDIA says it also publishes datasets and recipes. Some training inputs are private third-party datasets listed on the card, so the data side is open only in part.
Does Nemotron 3.5 Bolt learn from my prompts?
No. NVIDIA's privacy subcard says no user interaction data and no personal data were used to train it, and NVIDIA operates no hosted service for it. When you run it locally, your prompts go only where your agent app sends them, so check that app and any tools it calls.
Who should skip Nemotron 3.5 Bolt?
Skip it if you need the quickest local decoding, since Gemma 4 26B-A4B and Qwen3.6 35B-A3B both outpace it. Skip it for closed-book fact recall as well, where its AA-Omniscience accuracy is 16.90. Teams that want a managed API should look at Nemotron 3.5 Lightning or Super VL, which have hosted listings.
Top Alternatives
- Nemotron 3.5 Lightning: Pick Bolt if you need image and video input and stronger browsing scores; pick Lightning if you want a text-only model with hosted API options.
- Qwen3.8-27B: Pick Bolt if laptop decode speed matters, since NVIDIA measured it roughly three times faster; pick Qwen3.8-27B if raw scores matter more, as its HokAI record lists 89.2 on GPQA Diamond.
- Nemotron 3 Nano 30B A3B: Pick Bolt for vision input and newer agent training; pick Nano 30B A3B if you want the smaller, older model in the same family.