NVIDIA NIM is NVIDIA's endpoint at integrate.api.nvidia.com, the reference deployment for the Nemotron family, and it lists most of its catalog at $0 per Mtok in the registry Atlas reads. GLM-5.2, MiniMax-M3, Kimi K2.6, Gemma 4 31B, and GPT OSS 120B are all listed at $0/$0. The paid tiers are Nemotron 3 Ultra at $0.50/$2.50, Nemotron 3 Super at $0.20/$0.80, and DeepSeek V4 Pro at $0.435/$0.87. NVIDIA NIM is the cheapest legitimate way to trial big open weights before you commit to a paid host, but the $0 endpoints are evaluation-tier and rate limited.
Are NVIDIA NIM models really free to use with Atlas?
Most NVIDIA NIM endpoints are listed at $0/$0 per Mtok in the registry Atlas reads, including GLM-5.2, MiniMax-M3, Kimi K2.6, Gemma 4 31B, and GPT OSS 120B. The catch is that those $0 endpoints are evaluation-tier and rate limited, so they are not a production SLA.
The $0 listing on NVIDIA NIM is real and it is also not what it looks like. As a way to try a model before you pay for it, NVIDIA NIM is unmatched: GLM-5.2 at a 1,000,000 token window and Kimi K2.6 at 262,144 are both listed at $0/$0, so you can benchmark Chinese frontier open weights on US-hosted NVIDIA infrastructure without a bill. What those endpoints will not do is carry a sustained Atlas agent loop, which hammers a provider with re-prompted turns full of retrieved code and unified diffs. Sustained agent loops need the paid Nemotron tiers or a different host. Treat NVIDIA NIM's free tier as a benchmark harness, not as production capacity.
Where is Nemotron 3 Ultra cheapest to run?
NVIDIA NIM serves Nemotron 3 Ultra 550B A55B at a 1,000,000 token context for $0.50/$2.50 per Mtok, the lowest listed price for that model across every host in the registry. NVIDIA NIM is the reference deployment for the Nemotron family, which is why the home-turf pricing holds.
Buying a model from the lab that made it is usually the cheapest path, and Nemotron on NVIDIA NIM is a clean example. Nemotron 3 Ultra 550B A55B at $0.50/$2.50 per Mtok with a 1,000,000 token window is a lot of reasoning capacity for the price, and Nemotron 3 Super at $0.20/$0.80 is the cheaper step down. Pin "model": "nvidia/nvidia/nemotron-3-ultra-550b-a55b" in atlas.json for the 1M-context reasoning slot. The same pattern shows on DeepSeek: DeepSeek V4 Pro on NIM is $0.435/$0.87 per Mtok versus $1.74/$3.48 on Together AI and Fireworks, a 4x input spread for identical weights.
Why does atlas models nvidia return so much noise?
In 2026 the NVIDIA NIM catalog is padded with non-coding artifacts: rerankers, PII taggers, content-safety classifiers, FLUX image models, and autonomous-driving nets. Running `atlas models nvidia` returns a lot of noise, so filter for the nemotron rows, which are the ones NVIDIA actually tunes for tool use.
NVIDIA NIM is an inference platform for NVIDIA's whole model portfolio, not a coding-agent catalog, so the listing Atlas resolves from models.dev includes a great deal that has nothing to do with writing code. The workflow that cuts through it is simple: run `atlas models nvidia` and filter for the nemotron rows, which are the ones NVIDIA actually tunes for tool use. That matters because every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, and a model that was tuned as a content-safety classifier will not produce the tool calls that gate expects. Then use `/models` in the TUI to add your pick to favorites.
How do you configure NVIDIA NIM in Atlas?
Get an API key from build.nvidia.com and export it as NVIDIA_API_KEY. Atlas resolves NVIDIA from models.dev with @ai-sdk/openai-compatible against https://integrate.api.nvidia.com/v1, so no custom provider block is needed in atlas.json for NVIDIA NIM.
Configuration for NVIDIA NIM is the shortest of any provider because Atlas already knows the endpoint. Export NVIDIA_API_KEY, and the registry does the rest through @ai-sdk/openai-compatible against https://integrate.api.nvidia.com/v1. From there, pin the model you want in atlas.json and use `/models` in the TUI to add it to favorites so `model.cycle_recent` can flip between NIM and your frontier provider. That flip is the practical shape of an NVIDIA NIM setup, because Atlas lets you switch the active model and provider on the fly with favorites and recents, and NIM's evaluation-tier endpoints pair naturally with a paid provider that catches the turns NIM rate limits.
When should you pick a different provider than NVIDIA NIM?
Pick a different provider than NVIDIA NIM for sustained production load. The $0 endpoints on NVIDIA NIM are evaluation-tier and rate limited, so they are not a production SLA, and sustained agent loops need the paid Nemotron tiers or a different host entirely.
NVIDIA NIM is best understood as two products sharing an endpoint. The free half, with GLM-5.2, MiniMax-M3, Kimi K2.6, Gemma 4 31B, and GPT OSS 120B listed at $0/$0, is a trial surface: use it to decide, not to ship. The paid half, Nemotron 3 Ultra at $0.50/$2.50, Nemotron 3 Super at $0.20/$0.80, and DeepSeek V4 Pro at $0.435/$0.87, is genuinely competitive and in Nemotron's case is the cheapest listing in the registry. If neither half fits, a host with a broader open-weights catalog is the alternative, and Atlas makes switching a one-line change in atlas.json rather than a rewrite.
Setup
- 01Get an API key from build.nvidia.com and export it as NVIDIA_API_KEY
- 02Atlas resolves NVIDIA from models.dev with @ai-sdk/openai-compatible against https://integrate.api.nvidia.com/v1, so no custom provider block is needed
- 03Run `atlas models nvidia` and filter for the nemotron rows, which are the ones NVIDIA actually tunes for tool use
- 04Pin `"model": "nvidia/nvidia/nemotron-3-ultra-550b-a55b"` in atlas.json for the 1M-context reasoning slot
- 05Use `/models` in the TUI to add it to favorites so `model.cycle_recent` can flip between NIM and your frontier provider
Frequently asked questions
- is nvidia nim free to use
- Most NVIDIA NIM endpoints are listed at $0/$0 per Mtok, including GLM-5.2, MiniMax-M3, Kimi K2.6, Gemma 4 31B, and GPT OSS 120B. Those endpoints are evaluation-tier and rate limited, so they are not a production SLA.
- how do I set up NVIDIA_API_KEY for atlas
- Get an API key from build.nvidia.com and export it as NVIDIA_API_KEY. Atlas resolves NVIDIA from models.dev with @ai-sdk/openai-compatible against https://integrate.api.nvidia.com/v1, so no custom provider block is needed.
- cheapest place to run nemotron 3 ultra
- NVIDIA NIM, at $0.50/$2.50 per Mtok with a 1,000,000 token context. That is the lowest listed price for Nemotron 3 Ultra 550B A55B across every host in the registry Atlas reads.
- deepseek v4 pro pricing comparison across providers
- DeepSeek V4 Pro on NVIDIA NIM is $0.435/$0.87 per Mtok versus $1.74/$3.48 on Together AI and Fireworks, a 4x input spread for identical weights.
- why does atlas models nvidia show so many models
- The NVIDIA NIM catalog is padded with non-coding artifacts like rerankers, PII taggers, content-safety classifiers, FLUX image models, and autonomous-driving nets. Filter for the nemotron rows, which NVIDIA actually tunes for tool use.
- can I use nvidia nim in production
- The $0 endpoints are evaluation-tier and rate limited, so they are not a production SLA. Sustained agent loops need the paid Nemotron tiers, such as Nemotron 3 Super at $0.20/$0.80, or a different host.
- how do I switch between nvidia nim and another provider in atlas
- Use `/models` in the TUI to add the model to favorites so `model.cycle_recent` can flip between NIM and your frontier provider. Atlas lets you switch the active model and provider on the fly with favorites and recents.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs GitHub Copilot CLI: A Developer's Guide for 2026
Comparing Atlas and GitHub Copilot CLI for developers in 2026. Explore pricing, privacy, agentic workflows, and extensibility to choose the best terminal AI coding agent.
Atlas for Astro: Islands, Content Collections, and Zero JS by Default in 2026
Atlas is a terminal-native AI coding agent for Astro in 2026. It reads astro.config.mjs, src/pages, and content collection schemas, drops needless client:load directives, and runs astro check.
Atlas for COBOL: Copybooks, PIC Clauses, and GnuCOBOL in 2026
Atlas is a terminal-native AI coding agent for COBOL in 2026. It reads your divisions and copybooks, explains a paragraph before touching it, and compiles with cobc under GnuCOBOL.
Atlas vs Qwen Code: Terminal AI Coding Agents in 2026
Atlas and Qwen Code are leading terminal AI coding agents in 2026. Compare Atlas's TUI, permission-gated tools, and free core with Qwen Code's 1M token context, Agent Teams, and Alibaba Cloud hosting.
Atlas vs Amp: Terminal AI Coding Agents in 2026
Compare Atlas, a terminal-native AI coding agent with free core and local embeddings, against Amp, Sourcegraph's agent featuring Oracle and Orbs, for developers in 2026.
Atlas for Ruby on Rails in 2026
Atlas is a terminal-native AI coding agent for Ruby on Rails in 2026. Run it in a Rails app with a config/application.rb and review every diff before it lands.
Atlas for Electron: Terminal-Native AI Coding for Main, Preload, and Renderer in 2026
Atlas is a terminal-native AI coding agent for Electron in 2026, where the main and renderer split, contextIsolation, and preload bridges are the security model.
Atlas for Next.js in 2026
Adopt Atlas, the terminal-native AI coding agent, for Next.js development in 2026. Enhance productivity across App Router, server components, and API routes with secure, reviewable AI assistance.