You've probably seen the post: "Stop paying $20/month, NVIDIA hosts 80+ AI models for free." It's been reshared everywhere since June. The claim is true. I hit the endpoint this morning. What the post leaves out: a terms-of-service document...
You've probably seen the post: "Stop paying $20/month, NVIDIA hosts 80+ AI models for free." It's been reshared everywhere since June.
The claim is true. I hit the endpoint this morning. What the post leaves out: a terms-of-service document that says "not in production" twice, a clause that lets NVIDIA use your prompts to improve its models, a rate limit that NVIDIA's own moderators say will never be raised, and the fact that three of the five models the original post named have already been removed.
I spent two days on this: the trial terms PDF, four months of developer-forum threads, the live model endpoint, and the leaderboards. Here's what "free" actually means, with sources, and the setup that works inside the limits.
What NVIDIA Build is, in one paragraph
Build (build.nvidia.com) is the catalog. NIM (NVIDIA Inference Microservices) is the packaging: a model in a container tuned for NVIDIA hardware, sold to enterprises. The hosted versions of those containers are what you call for free. Endpoint https://integrate.api.nvidia.com/v1, key prefix nvapi-, wire format identical to OpenAI chat completions. Free tier since July 2024 for anyone in the NVIDIA Developer Program, which is also free. No card.
Why give it away? Because the hosted endpoint is a showroom. The product is the container you run yourself under an NVIDIA AI Enterprise license, list price $4,500 per GPU per year. Every limit below follows from that.
80 models. Here's what they actually are.
Every post gives a different count: 80+, 100+, 160+. The catalog changes weekly, and the website and the API disagree.
The API is ground truth, and it answers without a key:
curl -s https://integrate.api.nvidia.com/v1/models | jq '.data | length'
# 80 (15 Oct 2026)
16 are current-generation chat or reasoning models. These are what you came for.
32 of the 80 are NVIDIA's own, including Nemotron 3 Ultra and Super (strong) and a long tail of safety classifiers, embedders and a 340B reward model you'll never call from a chat client.
26 are general chat models from 2023 to 2025: Llama 2 70B, Mixtral 8x22B, Yi Large, DBRX. They work. Nobody should pick them today.
The sixteen worth your time, with ids as the API expects them:
Model id
Context
Note
z-ai/glm-5.3
1M
#2 open model on the AA index
z-ai/glm-5.3-flash
1M
#4, fast
deepseek-ai/deepseek-v4.1-flash
1M
#5, 174 tok/s elsewhere
moonshotai/kimi-k2.6
262K
Kimi K3 is on the site, not in the anonymous list (see below)
nvidia/nemotron-3-ultra-550b-a55b
1M
NVIDIA flagship, 55B active
nvidia/nemotron-3-super-120b-a12b
262K
google/gemma-4-31b-it
262K
meta/muse-glimmer-30b
131K
openai/gpt-oss-20b
131K
the 120B is gone
mistralai/mistral-large
stepfun-ai/step-5-preview
new this month
poolside/laguna-xs-2.1
262K
coding
3 small Nemotron 3/3.5 variants, DiffusionGemma
cheap, fast
Are they any good? Better than the headline suggests.
"80 free models" is the wrong number to care about. The right question is whether the ones that matter are there. I checked the catalog against the Artificial Analysis open-weights index.
Six of the eleven highest-ranked open-weight models in the world are on this endpoint for nothing. GLM-5.3 ranks second overall. Elsewhere it costs about $0.90 per million output tokens; DeepSeek V4.1 Flash about $0.20. Vendor-reported: Kimi K3 scores 93.5 on GPQA Diamond and 76.8 on SWE-bench Verified; DeepSeek V4.1 Flash reports 90.6 on Terminal-Bench 2.1. These are in the same conversation as the proprietary models people pay $20/month for.
Two asterisks. Kimi K3 is on the website's featured list but wasn't in the /v1/models response without a key; some models need you to click "Try API" on their page once before the standard key works. And NVIDIA's own Nemotron 3 Ultra sits tenth of eleven on that index. NVIDIA is hosting its competitors' better models for free. Credit for that.
Catch 1: 40 requests per minute, and no way up
NVIDIA doesn't publish the limit. The ~40 RPM per key, shared across all models, comes from thousands of developers hitting the same wall. Staff confirmed on the forum in September 2025 that the old credit system "has been replaced by rate limits for trial usage," that limits "vary by model and are not published," and that you can see your own ceiling top-right on build.nvidia.com after signing in.
Forty a minute is plenty for a human in a chat window. It is nothing for an agent.
The developer forum has at least seven threads titled some variant of "Rate limit increase, 40 → 200 RPM," nearly all from people pointing coding agents at it: OpenCode, Cline, Claude Code. One describes 429s "repeatedly blocking agent workflows" even with exponential backoff. The moderator reply, quoted across those threads:
"There is no official way to circumvent this rate limit or to receive a rate limit increase on that same tier."
And the recommended path if you need more: "pay and deploy a model through NVIDIA NIM / NVIDIA Build." The showroom, doing its job.
What this means in practice: great for one chat window, a single-threaded coding agent with parallel tool calls off, batch scripts with a sleep, and A/B-testing models. Wrong backend for anything with fan-out.
Catch 2: "not in production," twice, and your prompts may train NVIDIA's models
I read the NVIDIA API Trial Terms of Service (October 2025 revision) clause by clause. As far as I can tell, nobody else writing about this has.
Three clauses deserve more than a table row.
Production is out, in two places. Clause 1.2: access is "for limited trial purposes only and without use of the API Service or Generated Content in production." Clause 1.4 repeats it: "you may only use the API Service for internal testing and evaluation purposes, not in production." Staff defined trial use as "prototyping, research, development, testing, or learning," and added that "the trial period is not limited by a time period." Your side project is fine forever. The day it has users, it isn't.
Your data has two clauses that look contradictory. 2.3: NVIDIA "will not store or use User Content or Generated Content at the end of each API Service session." Good. But 2.3 opens with "except as stated below in Section 2.4 or unless expressly disclosed," and 3.3(iv) says NVIDIA collects, "without identifying specific users," your "User Content and Generated Content to improve NVIDIA products and services, including AI models." My reading: NVIDIA doesn't keep your session as your session, but can keep what you sent, de-identified, to improve models. When users asked staff directly whether prompts are logged or used for training, the answer pointed back to the terms and did not say no. Treat it as: anything you send may end up in a training set.
Two restrictions aimed at developers. 4.10: no security testing of the service, including "probe, scan, or vulnerability test." 4.12: you won't use the service "including Generated Content to develop or improve products or services that compete with the API Service." If the plan was to distil GLM-5.3 outputs into your own hosted model, read that one with a lawyer.
On the plus side, 6.3: you own the outputs, subject to each model's license. GLM and DeepSeek are MIT. Nemotron is under the NVIDIA Open Model License.
Catch 3: models vanish without notice
The post that started this was published 14 June 2026. It named MiniMax, Kimi, DeepSeek, GLM and GPT-OSS. Four months later:
MiniMax: gone from the catalog.
GPT-OSS-120B: gone. Only the 20B remains.
DeepSeek V4 Pro and the original V4 Flash: deprecated 7 August 2026. A user asked why the same day. No official answer. Another user's guess, that capacity was freed for Kimi K3 and V4.1 Flash, is plausible and unconfirmed. "A seven-day notice would have been nice." There was none.
Not malice. The economics of a free showroom: finite GPUs, new models monthly, no SLA. Clause 1.3 says it outright: NVIDIA "may also choose to abandon development and terminate the availability of a pre-release service at any time without liability."
Rule: never hard-code a model id from a blog post, including this one. Query /v1/models at startup and have a fallback.
So does it replace $20 a month?
For most ChatGPT Plus or Claude Pro subscribers: no, and the comparison is a little dishonest. You're not paying $20 for a model. You're paying for an app with memory, voice, file upload, image generation, a mobile client and uptime. The free endpoint is an API.
For developers the answer flips. If you live in a terminal and an editor, the free endpoint plus Open WebUI gives you a ChatGPT-style interface to frontier-class open models for nothing, with the three catches above. For evaluating which open model to build on before you pay anyone, nothing else comes close.
Versus the other free tiers (October 2026)
The landscape has thinned. GitHub Models was retired on 30 July 2026. Cerebras replaced its free tier with a $5 credit that expires in 30 days and needs a card. Groq ended Llama 3.3 70B in June. What's left:
Provider
RPM
Per day
Card
Data caveat
NVIDIA NIM
~40
~10,000 (community-reported)
No
Content may improve NVIDIA models (3.3 iv)
Groq
30
1,000 (8K TPM binds first)
No
None stated
OpenRouter :free
20
50, or 1,000 after $10 top-up
No
Free providers may log
Google AI Studio
5–15
~1,500
No
Trains on your data outside EEA/UK/CH, human review
Cloudflare Workers AI
10k neurons/day
No
Mistral Experiment
~$10 credit/mo
No
Trains unless opted out
Cerebras
$5 credit, 30 days
Yes
GitHub Models
retired 30 Jul 2026
Five of OpenRouter's 14 current :free models are NVIDIA Nemotron variants. Same subsidy, different door.
The setup several people have converged on: chain them. NVIDIA first for headroom, Groq second for speed, OpenRouter last as the spare.
Setup that works inside the limits
Get the key: sign in at build.nvidia.com, open any model page, "Get API Key." Shown once.
curl https://integrate.api.nvidia.com/v1/chat/completions \
-H "Authorization: Bearer $NVIDIA_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"z-ai/glm-5.3",
"messages":[{"role":"user","content":"Explain NIM in one sentence."}],
"max_tokens":200}'
The fallback chain. The line that matters is max_retries=0: the OpenAI client's default retry sleeps on a 429, which is the last thing you want when the next provider is one line away.
import os
from openai import OpenAI, RateLimitError
CHAIN = [
("nvidia", "https://integrate.api.nvidia.com/v1", os.environ["NVIDIA_API_KEY"], "z-ai/glm-5.3"),
("groq", "https://api.groq.com/openai/v1", os.environ["GROQ_API_KEY"], "openai/gpt-oss-120b"),
("openrouter", "https://openrouter.ai/api/v1", os.environ["OPENROUTER_API_KEY"], "nvidia/nemotron-3-super-120b-a12b:free"),
]
def ask(prompt):
for name, base, key, model in CHAIN:
client = OpenAI(base_url=base, api_key=key, max_retries=0)
try:
r = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": prompt}],
max_tokens=400)
return name, r.choices[0].message.content
except RateLimitError:
continue
raise RuntimeError("every free tier is rate-limited right now")
provider, answer = ask("What changed in GLM-5.3?")
print(provider, "→", answer)
Open WebUI, if you want the chat-app experience:
docker run -d -p 3000:8080 \
-e OPENAI_API_BASE_URL=https://integrate.api.nvidia.com/v1 \
-e OPENAI_API_KEY=$NVIDIA_API_KEY \
-v open-webui:/app/backend/data \
--name open-webui ghcr.io/open-webui/open-webui:main
Open localhost:3000, pick z-ai/glm-5.3, and you have what the $20 subscription sells, minus voice and image generation, at 40 messages a minute.
Tool-specific gotchas, briefly: Cline/Roo: turn off parallel tool calls. aider: prefix the model with openai/ so LiteLLM routes it. Cursor: features that go through Cursor's own backend ignore your base-URL override; chat works, the rest may not.
What production actually costs
The irony: once you know which open model you want, paying per token for it through any commercial host is usually cheaper than the $20 you were avoiding. GLM-5.3 at $0.90 per million output tokens means ten million output tokens, which is a lot of chatting, costs nine dollars. The free tier's real value isn't saving $20. It's letting you find that out before spending anything.
The enterprise rung, $4,500 per GPU per year at list, is for teams who want the containers on their own hardware. Quieter and more useful: NVIDIA lets you self-host NIM containers free for research and development on up to 16 GPUs, if you already own them.
Who should use this
Use it if you're evaluating open models, prototyping, learning, running a personal chat UI, or comparing GLM-5.3 vs Kimi K3 vs DeepSeek before picking one to pay for.
Use it if you run a single-threaded coding agent and can live with occasional 429s.
Don't if the thing calling it has users, handles customer data, fans out to parallel agents, or needs the same model id tomorrow.
Don't if "free" was the only reason. Nine dollars of pay-per-token buys the same models with none of the catches.
Sources
NVIDIA API Trial Terms of Service, PDF dated 7 Oct 2025 (assets.ngc.nvidia.com/products/api-catalog/legal/)
GET https://integrate.api.nvidia.com/v1/models, 15 Oct 2026, 80 ids
NVIDIA Developer Forums: "'Request More' (+4,000 credits) option" (staff, 10 & 17 Sep 2025); "NIM rate limit" (Jun 2026); the 40→200 RPM threads (Jun–Aug 2026); "Model Deprecation Request" (Jul–Aug 2026); "Clarification On Trial API Use" (May–Jun 2025)
Artificial Analysis open-weights Intelligence Index v4.3.2, read 15 Oct 2026
NVIDIA AI Enterprise licensing guide, docs.nvidia.com
OpenRouter limits docs and models API, 15 Oct 2026; Google Gemini API Additional Terms ("Unpaid Services"); Cerebras pricing; GitHub Models retirement notice; T. Nasser, "Free LLM API Tiers in October 2026," dev.to, 7 Oct 2026
L. Zarantonello, "Stop Paying $20/month. Use NVIDIA Build," Medium, 14 Jun 2026
Everything above was true on 15 October 2026. Given how this catalog behaves, check /v1/models before trusting any of it next month.
The thing I'd like to hear from you: if you've hit the 40 RPM wall with a specific agent and found settings that keep it under, what were they? That's the part I have the least data on, and it's the difference between "prototyping sandbox" and "daily driver" for a lot of people.
If this was useful, follow for the next one. I'm going to get a key and measure the six top models properly: latency, tokens per second, tool calling, and exactly when the 429 arrives.