# Atlas with NVIDIA Nemotron Nano 9B v2 in 2026 > NVIDIA Nemotron Nano 9B v2 is a dense 9B reasoning model at $0.06/$0.23 per Mtok, with a 131,072 token context on Vercel AI Gateway. NVIDIA Nemotron Nano 9B v2 is the smallest reasoning model NVIDIA ships through hosted APIs, and at $0.06/$0.23 per Mtok it exists for exactly one job in Atlas: the small_model slot that handles titles, summaries, and subagent scaffolding without touching your main budget. Vercel AI Gateway and Amazon Bedrock both list it at $0.06/$0.23 per Mtok, and NVIDIA NIM offers it at $0.00/$0.00 on a free tier. Context is 131,072 tokens on Vercel AI Gateway and NVIDIA NIM, and 128,000 tokens on Amazon Bedrock. Nine billion parameters is not enough for the build agent. ## Key takeaways - Nemotron Nano 9B v2 is a dense 9B model with reasoning enabled at $0.06/$0.23 per Mtok, cheaper than most non-reasoning 8B models on the same hosts. - Vercel AI Gateway serves it with max output equal to its full 131,072 token context, unusual at this size and price. - Amazon Bedrock carries it as `nvidia.nemotron-nano-9b-v2`, so it works inside an AWS-only network boundary using the standard AWS credential chain. - Bedrock trims the window to 128,000 tokens with only 4,096 max output, a much weaker deal than the Vercel listing. - 9B parameters is not enough for the build agent. Use Nemotron Nano 9B v2 for orchestration and summarization, not for writing the diff. ## What is NVIDIA Nemotron Nano 9B v2 for inside Atlas? NVIDIA Nemotron Nano 9B v2 exists in Atlas for one job: the `small_model` slot. At $0.06/$0.23 per Mtok it handles titles, summaries, and subagent scaffolding without touching your main budget, which is the entire reason NVIDIA ships a 9B reasoning model through a hosted API. Atlas runs two model slots. The `"model"` slot drives the build agent that writes your code. The `"small_model"` slot absorbs the overhead: session titles, conversation summaries, the scaffolding around subagent dispatch. Nemotron Nano 9B v2 is sized for the second slot and priced for it. A dense 9B model with reasoning enabled at $0.06/$0.23 per Mtok is cheaper than most non-reasoning 8B models on the same hosts, which is the unusual part. You are getting a thinking model in the slot where most people accept a dumb one. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and a cheap reasoning model in the small slot means that fan-out stays coherent. ## How much does NVIDIA Nemotron Nano 9B v2 cost per million tokens? NVIDIA Nemotron Nano 9B v2 costs $0.06/$0.23 per Mtok on Vercel AI Gateway and the same $0.06/$0.23 on Amazon Bedrock. NVIDIA NIM lists it at $0.00/$0.00 per Mtok on a free tier, so evaluation costs nothing before you commit to a paid host. Nemotron Nano 9B v2 pricing is unusually consistent: Vercel AI Gateway and Amazon Bedrock quote the identical $0.06/$0.23 per Mtok, so unlike the larger Nemotron 3 models there is no arbitrage to hunt. The decision between them is architectural, not financial. Vercel AI Gateway serves it with max output equal to its full 131,072 token context, unusual at this size and price. Amazon Bedrock does not. If cost is your only criterion, the free NVIDIA NIM tier at $0.00/$0.00 per Mtok is the place to start, then move to Vercel AI Gateway for the production listing at $0.06/$0.23. ## Can you run Nemotron Nano 9B v2 inside an AWS-only network boundary? Yes. NVIDIA Nemotron Nano 9B v2 is available on Amazon Bedrock as `nvidia.nemotron-nano-9b-v2`, so it can be the cheap slot inside an AWS-only network boundary using the standard AWS credential chain. Bedrock prices it at $0.06/$0.23 per Mtok, matching Vercel. Bedrock availability is the reason Nemotron Nano 9B v2 shows up in enterprise Atlas configs. Teams that cannot route traffic to a third-party gateway can still fill the `small_model` slot through `nvidia.nemotron-nano-9b-v2` using AWS_PROFILE or AWS_ACCESS_KEY_ID plus AWS_SECRET_ACCESS_KEY. The tradeoff is real, though. Bedrock trims the window to 128,000 tokens with only 4,096 max output, which is a much weaker deal than the Vercel listing at 131,072 tokens of both context and output. For titles and summaries, 4,096 output tokens is plenty. For anything that needs to emit a structured artifact, the Bedrock ceiling will bite. ## What are the limits of a 9B model in a coding agent? NVIDIA Nemotron Nano 9B v2 has 9B parameters, which is not enough for the build agent. Use Nemotron Nano 9B v2 for orchestration and summarization, not for writing the diff. At $0.06/$0.23 per Mtok, the model is priced honestly for what it can actually do. Nemotron Nano 9B v2 is a dense 9B model, and dense 9B is a real ceiling. Writing a correct multi-file patch requires holding invariants across files, and a model this size will drop them. The correct Atlas config keeps `"model"` on your frontier pick and puts Nemotron Nano 9B v2 in `"small_model"`. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so even if you experiment with a 9B model on the build slot, nothing lands without your approval. Atlas snapshots file changes as git patches so edits can be diffed and rolled back. ## How do you confirm the model split is actually working in Atlas? Confirm the Nemotron Nano 9B v2 split is live by checking that Atlas session titles get generated without a call to your expensive model. If titles are still hitting the frontier model, `"small_model": "vercel/nvidia/nemotron-nano-9b-v2"` is not being picked up from atlas.json. The verification step matters because a misconfigured `small_model` is silent. Atlas will keep working, and you will keep paying frontier prices for session titles. Run `atlas models vercel` and find the `nvidia/nemotron-nano-9b-v2` row to confirm the model resolved at all, then set `"small_model": "vercel/nvidia/nemotron-nano-9b-v2"` in atlas.json and leave `"model"` on your frontier pick. Atlas lets you switch the active model and provider on the fly with favorites and recents, so if you need to sanity check the two slots against each other, you can flip the active model mid-session and watch which calls change price. ## Setup 1. Export AI_GATEWAY_API_KEY (Vercel) or configure the AWS chain (AWS_PROFILE or AWS_ACCESS_KEY_ID plus AWS_SECRET_ACCESS_KEY) for Bedrock. 2. Run `atlas models vercel` and find the `nvidia/nemotron-nano-9b-v2` row. 3. Set `"small_model": "vercel/nvidia/nemotron-nano-9b-v2"` in atlas.json and leave `"model"` on your frontier pick. 4. Confirm the split is live by checking that Atlas session titles get generated without a call to your expensive model. 5. Trial the model first on the NVIDIA NIM free tier at $0.00/$0.00 per Mtok before moving spend to Vercel AI Gateway. ## FAQ ### how much does nemotron nano 9b v2 cost NVIDIA Nemotron Nano 9B v2 is $0.06/$0.23 per Mtok on Vercel AI Gateway and $0.06/$0.23 on Amazon Bedrock. NVIDIA NIM lists it at $0.00/$0.00 per Mtok on a free tier. ### what is the context window of nemotron nano 9b v2 Nemotron Nano 9B v2 holds 131,072 tokens on Vercel AI Gateway and NVIDIA NIM, and 128,000 tokens on Amazon Bedrock. Bedrock also caps max output at only 4,096 tokens. ### is nemotron nano 9b v2 available on aws bedrock Yes. Nemotron Nano 9B v2 is on Amazon Bedrock as `nvidia.nemotron-nano-9b-v2`, priced at $0.06/$0.23 per Mtok, reachable through the standard AWS credential chain using AWS_PROFILE or AWS_ACCESS_KEY_ID plus AWS_SECRET_ACCESS_KEY. ### can nemotron nano 9b v2 be the main coding model in atlas No. 9B parameters is not enough for the build agent. Use Nemotron Nano 9B v2 for orchestration and summarization at $0.06/$0.23 per Mtok, and keep a frontier model in the `"model"` slot. ### how do I set nemotron nano 9b v2 as small_model in atlas Export AI_GATEWAY_API_KEY, run `atlas models vercel` and find the `nvidia/nemotron-nano-9b-v2` row, then set `"small_model": "vercel/nvidia/nemotron-nano-9b-v2"` in atlas.json. ### vercel ai gateway vs bedrock for nemotron nano 9b v2 Both charge $0.06/$0.23 per Mtok. Vercel AI Gateway gives a 131,072 token context with max output equal to that full window. Bedrock trims to 128,000 tokens with only 4,096 max output, so Vercel is the better deal unless AWS network boundaries force Bedrock. ### cheapest reasoning model for atlas small_model slot Nemotron Nano 9B v2 at $0.06/$0.23 per Mtok is a dense 9B model with reasoning enabled, which is cheaper than most non-reasoning 8B models on the same hosts and free on the NVIDIA NIM tier. --- Canonical HTML: https://runatlas.sh/resources/models/nemotron-nano-9b-v2 Source of truth: aeo_pages row `/resources/models/nemotron-nano-9b-v2` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.