Mercury 3 and Nemotron 3.5: Fast, Cheap AI Models Explained

Quick answer

Mercury 3 (from the US startup Inception, launched 19 October 2026) is a paid, cloud-only AI model that writes text using diffusion instead of word by word. Inception says it runs at about 1,000 tokens per second and costs $0.10 per million input tokens during its launch offer. Nemotron 3.5 (from Nvidia, launched 15 October 2026) is a family of open-weight models: Super VL (120B) for company servers, plus two smaller "Bolt" models built to run on your own PC. You can download them, use them commercially and keep your data in-house.

1. Want the fastest, cheapest hosted API and don't mind sending data to a vendor? Try Mercury 3.
2. Want to self-host, fine-tune or keep data private? Look at Nemotron 3.5.

Our honesty note: SR Infobiz has not tested either model. Every number below comes from the company that made the model or from an independent tester, and we say which one each time.

Why these two models matter this month

Most AI news in 2026 has been about bigger and smarter models. October brought two releases that focus on something most builders care about more: speed and cost.

1. Inception wants to prove that a different way of generating text, diffusion, can match normal models on quality while being several times faster.
2. Nvidia wants to give companies and PC owners a strong model they can run on their own hardware without paying per token.

Both matter if you're an Indian startup watching your cloud bill, or a developer building chat, search or agent features where every second of waiting counts.

What is Mercury 3?

Mercury 3 is the new flagship model from Inception, a startup founded in 2024 by computer science professors Stefano Ermon, Aditya Grover and Volodymyr Kuleshov (SiliconANGLE, 20 Oct 2026). Inception announced it on its blog and through a Business Wire press release on 19 October 2026.

Mercury 3 at a glance (vendor claims)

All of these come from Inception's launch blog (19 Oct 2026) unless noted:

1. Speed: 1,068 tokens per second on Nvidia GPUs. Inception says that's "4–8x faster" than GPT-6 Luna and Claude Haiku 5.5. SiliconANGLE (20 Oct 2026) reports that in Inception's comparison, Haiku 5.5 reached about 240 tokens/sec and GPT-6 Luna about 130.
2. Context window: 1 million tokens, up from 260,000 in Mercury 2.5.
3. Size: not disclosed. Inception only says it is "more than 40x larger than Mercury 1".
4. Price: list price $0.20 per million input tokens and $0.70 per million output tokens. During launch it is 50% off at $0.10 input / $0.35 output. Cached input is 90% cheaper.
5. Free trial: 100 million free API tokens, according to the launch blog.
6. Where to get it: Inception API Platform, OpenRouter, Baseten and Vercel. It is OpenAI-API compatible, so existing code needs only small changes.

Benchmark claims

Inception says Mercury 3 is "on par or better" than GPT-6 Luna and Claude Haiku 5.5. According to SiliconANGLE (20 Oct 2026), which reported Inception's own benchmark table, Mercury 3 won 8 of 12 tests. Its biggest lead was coding: 51.1 on SWE-Bench Atlas, versus 33.3 for GPT-6 Luna and 32.3 for Haiku 5.5. It lost to Haiku 5.5 on DeepSWE (46.3 vs 51.6) and on one long-context test.

Two things to keep in mind:

1. These are vendor numbers. SiliconANGLE noted that no outside group had published its own results yet.
2. No independent speed or quality score yet. When we checked on 24 October 2026, Artificial Analysis had no Mercury 3 model page (the URL returned "not found"). OrcaRouter's blog (Oct 2026) made the same point: the 1,068 tokens/sec figure "has no second party behind it" yet.

On price, note that SiliconANGLE (20 Oct 2026) lists GPT-6 Luna and Haiku 5.5 at $0.10 input / $0.50 output for typical prompts. So Mercury 3's list price is actually higher on output; it's the launch discount that makes it cheaper for now. The pitch is speed, not just price.

Early customer reports (also vendor-supplied)

Inception's blog quotes early users. These are testimonials picked by the vendor, so read them as examples, not proof:

1. Octen AI (real-time search on smart glasses): median generation time of 0.9 seconds and 2.2 seconds end to end.
2. RevisionDojo (tutoring app): hints and marked answers "come back in about a second".
3. TinyFish AI (web agents): Mercury 3 completed 50% of 50 live Mind2Web navigation tasks.

What is a diffusion LLM? (Plain English)

Almost every chatbot you've used is autoregressive. It writes one token (roughly a word piece) at a time, left to right. To write token 500, it must first finish tokens 1 to 499. That's why long answers "type out" slowly.

A diffusion language model works more like an artist sketching. It starts with a rough draft of the whole answer at once, then refines every position in parallel over a few passes until the text is clean. SiliconANGLE (20 Oct 2026) describes it this way: a Mercury model "drafts the whole response up front, then refines every position in parallel over a few passes."

An easy way to picture it:

1. Autoregressive: a person typing a letter, one word after another.
2. Diffusion: a photo coming into focus. The whole picture is there from the start, just blurry, and each pass makes it sharper.

Why it can be faster: the GPU does a few big parallel passes instead of hundreds of small sequential steps. Inception co-founder Volodymyr Kuleshov wrote on Hacker News (19 Oct 2026) that as reasoning models "think longer per answer, the speed advantage over autoregressive models of the same size grows."

The big open question has been whether diffusion can get smart as it gets bigger. Inception says yes: it claims intelligence across standard benchmarks rose 3x from Mercury 1 (mini) to Mercury 3 (Inception blog, 19 Oct 2026). That is still a vendor claim until independent testers publish results.

What is Nemotron 3.5?

Nemotron 3.5 is Nvidia's newest family of open-weight models, announced by Kari Briski, Nvidia's VP of Open Models and Software, on the Nvidia blog on 15 October 2026. (Some news sites, such as TechBriefly, covered it on 20 October.) It is a normal autoregressive model, not diffusion, but it is built for speed using a hybrid Mamba-2 and Transformer design with Mixture-of-Experts, meaning only part of the model works on each token.

The three models

Sources: Hugging Face model cards for Super VL and Bolt 74B (Nvidia, Oct 2026); Nvidia blog (15 Oct 2026).

All of them take text, images and video as input and produce text (Hugging Face model cards).

Licence: genuinely open

Both model cards list the OpenMDW-1.1 licence and say the models are "ready for commercial or non-commercial use". The licence text on Hugging Face grants rights "free of charge... without restriction" and puts no restrictions on the outputs you generate. Your main duties are to keep the licence and notices if you redistribute the model. Nvidia's blog (15 Oct 2026) adds that it ships open weights, open datasets and training recipes. As always, have your lawyer read the licence before you build a business on it.

Independent results: Artificial Analysis

Here, unlike Mercury 3, we do have outside numbers. Artificial Analysis (model page, accessed 24 Oct 2026) reports for Super VL:

1. Intelligence Index: 26, versus a median of 8 for comparable open-weight models.
2. Output speed: 453.4 tokens/sec (median across providers) and 0.73 s time to first token.
3. Price (median across providers): about $0.23 per million input and $0.57 per million output tokens.
4. Verbose: it generated 170M tokens while running the index, against a median of 100M.

Artificial Analysis's launch article (Oct 2026) also says Super VL is now the highest-scoring US open-weight model, but that the US still trails the UAE (K2 Horizon, 31) and China (MiMo-V2.6-Pro, 46) on open-weight intelligence. It scored Bolt 74B-A9B at 21, only 1.5 points behind Nvidia's much larger Nemotron 3 Ultra.

A fair reading: Super VL is very strong for its size and openness, but it is not a frontier model. OrcaRouter (Oct 2026) also warns that "#1" rankings on the page are inside a filtered class, not overall.

Nvidia's own claims

These come from Nvidia and haven't been independently checked:

1. Super VL is "5x faster" (Nvidia blog headline, 15 Oct 2026).
2. Bolt 74B is "the #1 open model in its class on PinchBench, and the #2 open model for BrowseComp" (Nvidia blog).
3. After post-training, Super VL finishes some agent tasks faster than rival closed models (TechBriefly, 20 Oct 2026, reporting Nvidia's claims).
4. Named users include EY, Perplexity, Salesforce, CrowdStrike and Ramp (TechBriefly, 20 Oct 2026).

How to run Nemotron 3.5 locally

Bolt 74B-A9B 24GB is the one built for your own machine. Its weights are compressed to 2-bit so the whole model fits in 24GB of memory (Hugging Face model card). Nvidia's blog (15 Oct 2026) says it works with llama.cpp, Ollama, LM Studio, Unsloth and vLLM and runs on GeForce RTX PCs, DGX Spark, DGX Station, RTX PRO workstations and the new RTX Spark laptops.

A simple path:

1. Check your hardware. You'll need a GPU or unified-memory machine with roughly 24GB available for the model, plus room for everything else.
2. Install a local runner you already know, such as LM Studio or Ollama.
3. Search for "Nemotron 3.5 Bolt" in its model catalogue, or download the weights from Nvidia's Hugging Face page.
4. Point your local agent or app at it. Nothing leaves your machine.

Super VL is a server model. Its full-precision card lists 4× H100 or 2× GB200 GPUs for deployment. Most people will reach it through a cloud provider; Artificial Analysis lists several.

Mercury 3 vs Nemotron 3.5: comparison table

Who should use which?

Budget developers and hobbyists

Grab Mercury 3's 100 million free tokens and test it on your real prompts. The launch price is low, but plan for the list price ($0.70 per million output tokens) when the offer ends. If you have a strong gaming GPU or a Mac-class machine with enough memory, Bolt costs nothing per token.

Indian startups

1. Customer support bots, voice agents, tutoring and search where speed matters: Mercury 3 is worth a trial. Its customer stories are exactly these use cases.
2. Fintech, health, insurance or anything with sensitive customer data: Nemotron 3.5 lets you keep the model and data on your own servers, which helps when you have to answer questions about where data is stored. Nvidia's blog uses credit review and insurance claims as examples.
3. Cost control at scale: compare Mercury 3's per-token bill against renting GPUs for Super VL. Which is cheaper depends entirely on your volume.

Privacy-first and self-hosting teams

Choose Nemotron 3.5. You can't self-host Mercury 3. It's only offered as an API, though Inception says enterprise plans include configurable data retention. At SR Infobiz we build privacy-first tools that run in your browser, so we like that Bolt brings real agent capability to a local machine.

How to try them today

1. Mercury 3: sign up on Inception's API Platform (100M free tokens), try Mercury Chat, or pick it on OpenRouter, Baseten or Vercel.
2. Nemotron 3.5: download from Nvidia's Hugging Face page, run Bolt in LM Studio, Ollama or llama.cpp, or use Super VL through a hosted provider listed on Artificial Analysis.

Tip: test both with your own 20–50 real prompts and measure speed and answer quality yourself. Vendor benchmarks rarely match your workload.

FAQ

Is Mercury 3 open source?

No. It is available only through Inception's API and partner platforms (Inception blog, 19 Oct 2026).

Is Nemotron 3.5 free for commercial use?

Nvidia's Hugging Face model cards say it's ready for commercial use under the OpenMDW-1.1 licence, which carries no royalty. You still pay for the hardware or cloud you run it on.

Is a diffusion LLM less accurate?

Not necessarily. Inception claims Mercury 3 matches or beats speed-focused rivals on most of its tests, but it also lost some. Wait for independent results before you decide.

Can I run Nemotron 3.5 on my laptop?

The Bolt 74B model needs about 24GB of memory (Hugging Face model card). The smaller Bolt E4B (4GB) was announced for 23 October and is aimed at lighter devices (Nvidia blog, 15 Oct 2026).

Did SR Infobiz test these models?

No. We ran no tests of our own. This article summarises vendor announcements and independent results from Artificial Analysis, with sources below.

Sources

All accessed 24 October 2026.

1. Inception, "Diffusion at Scale: Introducing Mercury 3", 19 Oct 2026: https://www.inceptionlabs.ai/blog/introducing-mercury-3
2. Business Wire, "Inception Launches Mercury 3, Setting a New Pareto Frontier in Speed and Intelligence", 19 Oct 2026: https://www.businesswire.com/news/home/20261019090400/en/Inception-Launches-Mercury-3-Setting-a-New-Pareto-Frontier-in-Speed-and-Intelligence
3. SiliconANGLE (Duncan Riley), "Inception says Mercury 3 beats GPT-6 Luna and Claude Haiku 5.5 at up to 8x the speed", 20 Oct 2026: https://siliconangle.com/2026/10/20/inception-says-mercury-3-beats-gpt-6-luna-and-claude-haiku-5-5-at-up-to-8x-the-speed/
4. Hacker News launch thread, comment by Inception co-founder Volodymyr Kuleshov, 19 Oct 2026: https://www.betterhackernews.com/item/50155245
5. Nvidia Blog (Kari Briski), "Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models", 15 Oct 2026: https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/
6. Hugging Face, NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 model card: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16
7. Hugging Face, NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB model card: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
8. OpenMDW License Agreement v1.1 (on Hugging Face): https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr/blob/main/LICENSE
9. Artificial Analysis, Nemotron 3.5 Super VL model page: https://artificialanalysis.ai/models/nemotron-3-5-super-vl
10. Artificial Analysis, "NVIDIA Releases Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B A9B", Oct 2026: https://artificialanalysis.ai/articles/nemotron-3-5-super-and-bolt
11. TechBriefly (Aytun Çelebi), "Nvidia launches open-weight AI models for enterprise agents and PCs", 20 Oct 2026: https://techbriefly.com/2026/10/20/nvidia-nemotron-35-super-vl-open-weight-ai-models/
12. OrcaRouter, "Mercury 3 vs Nemotron 3.5 Super VL: Measured vs Claimed", Oct 2026: https://www.orcarouter.ai/blog/mercury-3-vs-nemotron-3-5-super-vl-120b