What NVIDIA says about the limit
In September 2025, an NVIDIA staff member wrote on the developer forum: "We no longer use a credit-based system for build.nvidia.com." Rate limits for trial usage replaced the credits. The same reply says that the limits vary for each model and that NVIDIA does not publish them. You can see your maximum rate limit at the top right of build.nvidia.com.
The NIM FAQ says that the number of calls allowed in a period "is designed to serve evaluation and prototyping needs only". Members of the NVIDIA Developer Program get free access to NIM API endpoints for prototyping.
Can you increase the 40 RPM limit?
The NVIDIA developer forum has many posts that ask for 200 RPM or more. We did not find an official process that raises the limit for a free account. One forum reply says that NVIDIA cannot manually increase limits for free personal accounts. We could not confirm that this reply is from NVIDIA staff.
The NIM FAQ gives other options for more capacity:
- Self-host NIM on your cloud provider or on your own GPUs.
- Use a production serverless NIM API from a partner, for example Hugging Face, with pay-per-use pricing.
- Start the 90-day NVIDIA AI Enterprise trial, which is meant for production deployments.
Do not make more NVIDIA accounts to get more requests. The trial terms let NVIDIA set use limits, and the free access is for limited trial purposes only.
| Option | Cost | More requests? |
|---|---|---|
| Ask on the forum for a higher free limit | Free | No official process found |
| Pace requests under the limit | Free | No, but fewer 429 errors |
| Add other free providers with failover | Free | Yes, from other providers |
| Serverless NIM from a partner | Pay per use | Yes |
| Self-host NIM | Your GPUs | Yes |
What the trial terms allow
Section 1.2 of the NVIDIA API Trial Terms gives access "for limited trial purposes only". It excludes use of the service or its output in production. Section 1.4 says that without a subscription, you may use the service only for internal testing and evaluation.
Section 3.3 says that NVIDIA collects your content and the generated content to improve its products, including AI models. Do not send private data to the trial service. Read free LLM API commercial use for other providers.
Free NIM models today
The catalog lists these free NVIDIA models. The limit is per account, so all models share the same requests per minute.
| Model | Free on | Context | Free limits (best offer) | Tools | Vision |
|---|---|---|---|---|---|
| GLM 5.3 Flash | NVIDIA NIM | 1M | 40 req/min 40 RPM | Yes | No |
| Step 5 Preview | NVIDIA NIM | - | Ongoing developer prototyping | No | No |
| Nemotron-3 Ultra 550B | NVIDIA NIM | 1M | 40 req/min 40 RPM | Yes | No |
| Muse Glimmer 30B | NVIDIA NIM | 131K | 40 req/min 40 RPM | Yes | Yes |
| Gemma 4 31B | NVIDIA NIM | 262K | 40 req/min 40 RPM | No | No |
| Poolside Laguna XS 2.1 | NVIDIA NIM | 262K | 40 req/min 40 RPM | Yes | No |
| Nemotron 3 Super 120B | NVIDIA NIM | 262K | 40 req/min 40 RPM | Yes | No |
| Nemotron 3.5 Lightning 30B A3B | NVIDIA NIM | 1M | 40 req/min 40 RPM | Yes | No |
| DiffusionGemma 26B A4B IT | NVIDIA NIM | 262K | 40 req/min 40 RPM | No | Yes |
| GPT-OSS 20B | NVIDIA NIM | 131K | 40 req/min 40 RPM | Yes | No |
| Nemotron 3 Nano Omni 30B A3B Reasoning | NVIDIA NIM | 262K | 40 req/min 40 RPM | No | Yes |
| Ising Calibration 1.5 31B | NVIDIA NIM | 131K | 40 req/min 40 RPM | No | Yes |
Showing 12 of 19 matching free models. See all models.
Live data from the FreeLLMAPI catalog (version 2026.11.19, checked 19 November 2026). Free limits change often. The catalog is checked every day.
| Provider | Known issue | What to do |
|---|---|---|
| NVIDIA NIM | Recurring free, 40 RPM, eval-only ToS | NVIDIA NIM replaced its depleting trial credits with a recurring per-account rate limit (40 RPM default, varies by model), verified June 2026. The trial ToS still scopes usage to evaluation/prototyping, not production. |
| NVIDIA NIM | Developer evaluation and data-use terms | Ongoing zero-price hosted developer access, not a monthly credit grant or unrestricted production service. NVIDIA trial/prototyping and data-use terms apply. Do not submit personal or confidential data. Text completions with exact model identity passed on October 15, 2026; tool and vision flags remain off until separately tested. Sources: https://build.nvidia.com/stepfun-ai/step-5-preview and https://openrouter.ai/nvidia/nemotron-3.5-super-vl:free |
Live data from the FreeLLMAPI catalog (version 2026.11.19, checked 19 November 2026). Free limits change often. The catalog is checked every day.
How to stay under the limit
At 40 requests per minute, you can send one request every 1.5 seconds on average. A burst of parallel requests reaches the limit in seconds. Use these rules:
- Limit the number of parallel requests in your client. Two or three at a time is safe for most agents.
- Retry each 429 with exponential backoff and jitter. Stop after a fixed number of retries.
- Count requests across all models. The limit is for the account, not for each model.
- Send short, simple tasks to other providers. Keep NVIDIA for the large models that only NVIDIA serves free.
Simple client-side pacing (Python)
import time
from openai import OpenAI, RateLimitError
client = OpenAI(base_url="https://integrate.api.nvidia.com/v1", api_key="<nvidia-key>")
MIN_GAP = 60 / 40 # seconds between requests at 40 RPM
last = 0.0
def ask(prompt, model="openai/gpt-oss-20b"):
global last
for attempt in range(5):
wait = MIN_GAP - (time.time() - last)
if wait > 0:
time.sleep(wait)
last = time.time()
try:
r = client.chat.completions.create(model=model, messages=[{"role": "user", "content": prompt}])
return r.choices[0].message.content
except RateLimitError:
time.sleep(2 ** attempt)
raise RuntimeError("NVIDIA rate limit: retries used up")Spread the load with FreeLLMAPI
FreeLLMAPI is an open-source gateway that runs on your computer. It treats NVIDIA as one pool of 40 requests per minute for the whole account. Before it sends a request, it checks that pool. When the pool is full, it sends the request to the next free provider in your chain.
- The account-wide minute cap is set with
PROVIDER_MINUTE_REQUEST_CAP_NVIDIA. Lower it if your account shows a smaller limit. - To limit parallel requests to NVIDIA, set
MAX_CONCURRENT_REQUESTS_PER_KEY_NVIDIA. - A 429 from NVIDIA puts the key on a short cooldown. The router then tries other providers.
- The
X-Routed-Viaheader shows which provider answered each request.
~/freellmapi/.env (Docker install)
PROVIDER_MINUTE_REQUEST_CAP_NVIDIA=30
MAX_CONCURRENT_REQUESTS_PER_KEY_NVIDIA=2
# then restart the container
cd ~/freellmapi && docker compose up -dThe one-line installer keeps its settings in ~/freellmapi/.env. Add the two lines to that file, then start the container again. The default NVIDIA cap is 40 requests per minute. A value of 0 turns the account-wide gate off for that provider.
Failover does not raise NVIDIA's limit. It uses other free tiers, for example Groq, Gemini and Cloudflare, when NVIDIA is full.
Frequently asked questions
What is the NVIDIA NIM API rate limit?
NVIDIA does not publish one number. Staff say that limits vary by model, and you can see yours at the top right of build.nvidia.com. Most free accounts report 40 requests per minute.
How do I increase the NVIDIA NIM rate limit from 40 RPM?
We found no official process for free accounts. NVIDIA's FAQ points to self-hosting, a paid serverless partner such as Hugging Face, or the NVIDIA AI Enterprise trial for more capacity.
Does NVIDIA NIM still use credits?
No. NVIDIA staff said in September 2025 that build.nvidia.com no longer uses credits. Trial use is now controlled by rate limits.
Can I use the free NVIDIA API in production?
No. The NVIDIA API Trial Terms allow limited trial use only, not production. Production use needs an NVIDIA AI Enterprise license or a paid provider.
Is the NVIDIA 40 RPM limit per model or per account?
Free users report one limit across the account. Plan for all models to share it, and pace requests across every model that you call.
Related guides
- free nvidia api
- Free LLM API rate limits
- Groq rate limit reached (429)
- Gemini API quota exceeded (429 RESOURCE_EXHAUSTED)
- Free LLM API for commercial use
Sources
- NVIDIA developer forum: credit system replaced by rate limits (checked 20 November 2026)
- NVIDIA NIM FAQ (checked 20 November 2026)
- NVIDIA API Trial Terms of Service (PDF) (checked 20 November 2026)
- NVIDIA developer forum: request to raise 40 RPM (checked 20 November 2026)
- NVIDIA API Catalog: API reference (checked 20 November 2026)
- FreeLLMAPI docs: Quota and cooldown engine (checked 20 November 2026)