On this page
nvidia/nemotron-3-super-120b-a12b (first token in 0.5 s, 136–200 tok/s in our runs); skip moonshotai/kimi-k3, which still does not answer. For raw speed, Groq is still several times quicker.NVIDIA NIM advertises the widest free catalogue we've tested: one key, dozens of models, no card. This is a hands-on review — we called the API, timed every response, and ran several of the models inside a real agent harness to see which ones can actually do work.
What you actually get
- 80 models listed, 13 that answered when we called each one on 11 October 2026 (81 and 12 on 29 September; 83 listed on 31 August). See the tables below.
- No credit card. Sign up at build.nvidia.com and generate a key starting
nvapi-. In 2026 this matters — half the "free" AI APIs we tested now demand a card. - OpenAI-compatible endpoint, so existing code works with a changed base URL.
- Limits are not published. Older guides (including an earlier version of this page) quote 1,000 signup credits, expandable to 5,000, and 40 requests/minute. NVIDIA's API docs state no credit count or rate limit as of 2026-10-11, and we have not hit a limit in our own runs, so treat those numbers as unconfirmed.
Which models actually answer (re-tested October 11, 2026)
We repeated the full sweep on 11 October 2026: GET /v1/models listed 80 models; we skipped 17 embedding, safety and reward models and sent each of the other 63 one short chat request, then timed the first token on a streaming call for every model that answered. 46 returned 404, 13 answered, poolside/laguna-xs-2.1 returned 503 ResourceExhausted (it answered in September), and three never answered, direct or through our proxy: moonshotai/kimi-k3, z-ai/glm-5.3 and meta/llama-3.2-90b-vision-instruct.
| Model ID that answered (2026-10-11) | First token | vs. 29 September |
|---|---|---|
| nvidia/nemotron-3-super-120b-a12b | 0.54 s | Still our default pick |
| nvidia/nemotron-3.5-lightning-30b-a3b | 0.64 s | Timed out in September; answers now |
| z-ai/glm-5.3-flash | 0.68 s | Timed out in September; answers now |
| meta/llama-3.2-11b-vision-instruct | 0.68 s | Still the only Llama chat model that answers |
| nvidia/riva-translate-4b-instruct-v2 | 0.69 s | Translation only |
| openai/gpt-oss-20b | 0.73 s | Reasoning model |
| nvidia/ising-calibration-1.5-31b | 0.76 s | Was 33 s in September |
| meta/muse-glimmer-30b | 1.30 s | |
| nvidia/nemotron-3-ultra-550b-a55b | 1.31 s | Largest Nemotron on the free key |
| deepseek-ai/deepseek-v4.1-flash | 1.38 s | Was 92 s in September; the only DeepSeek left |
| nvidia/nemotron-3-nano-omni-30b-a3b-reasoning | 1.45 s | Small reasoning model |
| google/diffusiongemma-26b-a4b-it | 8.41 s | Diffusion text model |
| google/gemma-4-31b-it | not timed | Answered on a retry; the timed call was dropped |
Also called on 11 October, from our older speed and agent tables: deepseek-ai/deepseek-v4-flash-0731 (end of life 2026-09-21), minimaxai/minimax-m3 (2026-09-09), mistralai/mistral-nemotron (2026-09-28) and nvidia/nemotron-3-nano-30b-a3b (2026-09-01) all return 410 Gone with those dates. First-token times are single calls from Asia, direct where the route worked; read them as rough.
The September 29 sweep
GET /v1/models returned 81 models. We skipped the 17 that are embedding, safety or reward models, then sent every remaining model one short chat request. 46 returned HTTP 404 "Function not found" — they are in the list but no longer served. These answered:
| Model ID | Time to answer "OK" | Notes |
|---|---|---|
| nvidia/nemotron-3-super-120b-a12b | 0.9 s | Our default pick for general use |
| nvidia/nemotron-3-ultra-550b-a55b | 0.9 s | Largest Nemotron on the free key |
| nvidia/nemotron-3-nano-omni-30b-a3b-reasoning | 1.6 s | Small reasoning model |
| openai/gpt-oss-20b | 1.0 s | Reasoning model; spends its first tokens thinking |
| meta/llama-3.2-11b-vision-instruct | 1.7 s | The only Meta Llama chat model that answered |
| meta/muse-glimmer-30b | 2.8 s | |
| poolside/laguna-xs-2.1 | 13.8 s | Coding model |
| nvidia/ising-calibration-1.5-31b | 33 s | Slow |
| google/gemma-4-31b-it | 56 s | Works, but queues |
| deepseek-ai/deepseek-v4.1-flash | 92 s | Works, but queues |
| google/diffusiongemma-26b-a4b-it | 0.7 s | Diffusion text model |
| nvidia/riva-translate-4b-instruct-v2 | 1.4 s | Translation only — it translated our prompt instead of answering |
Listed but no answer: moonshotai/kimi-k3 gave no response within 180 seconds, on two attempts with and without our proxy. z-ai/glm-5.3, z-ai/glm-5.3-flash and nvidia/nemotron-3.5-lightning-30b-a3b also timed out. Gone from the list since August: mistralai/mistral-nemotron and deepseek-ai/deepseek-v4-flash-0731, two of the four models in our speed table below.
The lesson is the same one we keep hitting across free APIs: a model appearing in /models does not mean you can call it. Send one tiny request to a model before you build on it.
Speed: we measured it
Re-measured 11 October 2026: nvidia/nemotron-3-super-120b-a12b generated 136.5 tok/s and 200.5 tok/s in two streaming runs (580 and 491 output tokens). The other three models below no longer answer — Mistral Nemotron and DeepSeek V4 Flash 0731 return 410 Gone, and Kimi K3 drops the connection. The original table, measured 31 August 2026:
Every model, same prompt, same US machine, streaming, timing generation only (excluding network latency) and using the API's own token counts:
| Model | Generation speed | Best for |
|---|---|---|
Kimi K3 moonshotai/kimi-k3 | 62.4 tok/s | No response within 3 min on 2026-09-29 |
Gemma 4 31B google/gemma-4-31b-it | 50.9 tok/s | Lightweight chat |
Mistral Nemotron mistralai/mistral-nemotron | 34.4 tok/s | General use |
DeepSeek V4 Flash deepseek-ai/deepseek-v4-flash-0731 | 27.3 tok/s | Reasoning, coding |
Two things to take from this. First, the spread within one platform is 2.3× — which model you pick matters more than which platform. Second, these August numbers were modest: Groq runs at roughly 350–840 tok/s (our 2026-10-11 re-test). Nemotron 3 Super's 136–200 tok/s in October narrows that gap, but NIM's selling point is still breadth, not speed.
The strange part: OpenRouter is faster at NVIDIA's own model
Update 2026-10-11: this has flipped. On the same prompt, NVIDIA's endpoint returned 491 tokens at 200.5 tok/s, while OpenRouter's free endpoint returned only about 36 visible tokens at 38.2 tok/s; a second attempt on each was cut off by our proxy. In August it was the other way round: we ran the identical model — nemotron-3-super-120b — on NVIDIA's endpoint and on OpenRouter's free tier. OpenRouter returned 45.7 tok/s. NVIDIA's own endpoint returned an empty response for the same request. Same weights, different serving infrastructure. If you specifically want Nemotron, it's worth testing both.
Which models can actually run an agent
Note (2026-10-11): these agent runs are from 31 August 2026 and were not repeated for this update. Of the models below, DeepSeek V4 Flash 0731 has since reached end of life (410) and Kimi K3 no longer answers; Nemotron 3 Super and Gemma 4 31B are still served.
Token speed says nothing about whether a model can complete a task. We plugged several NIM models into DeepSeek Harness and gave them real jobs — fix a broken Python script, tidy a folder, summarise CSVs — then checked the filesystem to see what actually happened:
| Model | Fix a broken script | Other tasks |
|---|---|---|
| Nemotron Super 120B | ✅ 43 s — fastest | ❌ Claimed a folder was tidied after 20 s; moved zero files |
| Kimi K3 | ✅ 56 s | ✅ Summarised three CSVs correctly in 64 s |
| DeepSeek V4 Flash | ✅ 251 s | — |
| Gemma 4 31B | ❌ Hung until 900 s timeout | — |
Note how badly throughput predicts this. DeepSeek V4 was the slowest model in the speed table yet completed the task; Gemma 4 was second fastest and never finished at all. And Nemotron's failure mode — reporting success without doing anything — is the one to watch for in any automation. Full method and results: we gave free AI models real agent jobs.
Working code
NIM is OpenAI-compatible, so only the base URL and key change:
from openai import OpenAI
client = OpenAI(
base_url="https://integrate.api.nvidia.com/v1",
api_key="nvapi-YOUR_API_KEY_HERE"
)
response = client.chat.completions.create(
model="nvidia/nemotron-3-super-120b-a12b", # first token in 0.54 s on 2026-10-11
messages=[{"role": "user", "content": "Explain GPU computing in one paragraph."}],
max_tokens=256,
)
print(response.choices[0].message.content)
Careful with older tutorials. Nearly every NIM guide online opens with model="meta/llama-3.3-70b-instruct". That model is gone — we hit a hard failure calling it in August 2026, and no Meta Llama chat models remain in the live catalogue. Substitute nvidia/nemotron-3-super-120b-a12b, which answered in under a second on 2026-09-29.
To see what's actually available on any given day, ask the API rather than trusting a blog post:
curl https://integrate.api.nvidia.com/v1/models \
-H "Authorization: Bearer $NVIDIA_API_KEY"
What it's genuinely good for
- Trying many models on one key. Few free keys give you DeepSeek, Kimi, Gemma, Mistral and Nemotron without five separate signups.
- Batch and offline work — summarising, classification, data extraction — where 30–60 tok/s is fine.
- Agent backends, with the caveat above: pick the model per job and verify the work got done.
- Beyond text: the catalogue also covers image generation, speech and vision models.
Where it falls short
- Not fast. For real-time chat or voice, Groq is roughly 10× quicker.
- Credits are finite (1,000–5,000, larger models cost more per call), so it's a place to experiment, not to run production traffic.
- The catalogue shifts. Llama's removal broke the code sample in every tutorial written before mid-2026 — including this one, until we retested.
- 40 requests/minute rules out heavy parallel fan-out.
Verdict
NIM is still a useful free catalogue — no card, and 13 working models behind one key (out of 80 listed on 2026-10-11), which is exactly what you want while you're still deciding which model fits your problem. Once you've decided, you'll likely move the workload somewhere faster or cheaper per token. Use it to explore; benchmark before you build on it.
All figures measured 2026-08-31 with our open harness: free-ai-api-benchmark. Free tiers change fast — we re-run rather than quote vendor pages.