Ray C Productions — Research Brief
Your key is live and working. Here's the honest picture — the good, the free, and the traps.
August 6, 2026 · Researched by 8 agents across NVIDIA docs, developer forums, GitHub and dev community · Every claim marked TESTED was run against your actual key today · Updated after benchmarking your own hardware — see section 04, which corrects an earlier claim
Bottom line: This is a genuine free buffet of ~100 AI models — no credit card, no bill. It is excellent for the small specialist jobs we don't do today (search matching, ranking, safety checks, images). It is bad as a replacement for anything already running in production.
Two hard limits decide everything: 40 requests per minute, and NVIDIA's own terms say the free tier is for prototyping and testing, not for shipping products.
Four names for what is basically one thing.
NVIDIA sells GPUs. To show off what their GPUs can do, they set up a giant free demo counter where they run about a hundred open AI models for anyone who signs up. That demo counter is what you now have a key to.
| Name you'll see | What it actually means |
|---|---|
| build.nvidia.com | The storefront. Where you browse models, click around, and get your key. |
| docs.api.nvidia.com | The instruction manual for calling those models from code. (The link you sent.) |
| NIM | NVIDIA's name for the box each model ships in. Marketing word — ignore it. |
| integrate.api.nvidia.com | The actual address our code talks to. |
NVIDIA copied OpenAI's exact interface. That means for most of our existing scripts, switching to NVIDIA is changing one line of text — the web address — not a rewrite. We already do this trick: JobHinge's classifier points at DeepSeek the same way.
Not marketing claims. Actual calls, actual clock times.
| What I asked for | Result | Speed | Notes |
|---|---|---|---|
| List available models | WORKS | fast | 102 models returned |
| Text embeddings nv-embedqa-e5-v5 | WORKS | 571 ms | 1024 dimensions — drop-in for job matching |
| Text embeddings nemotron-embed-1b-v2 | WORKS | 702 ms | 2048 dimensions, higher quality |
| Reranking nemotron-rerank-1b-v2 | WORKS | 308 ms | Scored a matching job +5.1, a bad match −12.5. Clean signal. |
| Image generation FLUX.1-dev | WORKS | 6.1 s | Returned a real image. Free. |
| Vision model llama-3.2-11b-vision | WORKS | 865 ms | Can read screenshots and photos |
| Safety filter nemoguard-content-safety | WORKS | 651 ms | Content moderation, free |
| Chat nemotron-super-49b | WORKS | 768 ms | Fast and good |
| Chat llama-3.3-70b | SLOW | 29 s | 29 seconds to say one word |
| Chat gpt-oss-120b | SLOW | 250 s | Over 4 minutes for a one-word answer |
| Chat deepseek-r1 | GONE | — | 404. Removed from catalog. |
| Chat qwen3-next-80b | GONE | — | Retired July 27, 2026 — 10 days ago |
| Embeddings nv-embedqa-1b-v2 | GONE | — | Retired May 18, 2026 |
| Reranking rerankqa-1b-v2 | GONE | — | Retired May 18, 2026 |
| Image FLUX.1-schnell | TIMEOUT | >60 s | Never answered |
| Image stable-diffusion-3 | BLOCKED | — | Not enabled for your account |
My very first call also came back "Worker limit reached (17/16)" — meaning I was queued behind other free users. That's the free tier in one sentence.
Nobody else is fighting over these, so there's no queue.
Developer forums are full of people hitting the 40/minute wall. NVIDIA staff have said flatly it won't be raised, and asking repeatedly may be treated as spam.
An earlier version of this page said Talos and Nami were too weak and NVIDIA would "fill the gap." I tested that claim. It was wrong.
I gave both engines the same job — pull the title, company, salary range and years of experience out of a job posting and return clean JSON — three times each, graded on exact correctness.
| Where | Model | Time to answer | Speed | Got it right |
|---|---|---|---|---|
| TALOS | llama3.2:3b | 1.5 s | 31 tok/s | 3 / 3 |
| TALOS | qwen2.5:3b | 1.9 s | 33 tok/s | 3 / 3 |
| TALOS | qwen2.5:7b | 3.6 s | 16 tok/s | 3 / 3 |
| NVIDIA | nemotron-3-nano-30b | 3.9 s | 71 tok/s | 3 / 3 |
| NVIDIA | nemotron-super-49b | 8.6 s | 28 tok/s | 3 / 3 |
| NVIDIA | nemotron-nano-9b | 5.0 s | 60 tok/s | 0 / 3 |
Five times faster, identical accuracy. The cloud models produce words 2–4× faster once they start — but you hand that back in network travel time and waiting in the free queue. Bigger did not mean better here.
Two other things this shook out:
A much narrower list than I first gave you — and notably, none of it is about raw horsepower:
1. Things you have no local equivalent for — image generation, quality reranking, quality embeddings. This is the real value.
2. Long documents — your local models have small memory; the hosted ones swallow far more text at once.
3. Genuinely hard writing and judgment — where a 3B model is shallow, not merely slower.
4. Reading images — screenshots, photos, scanned pages.
Fair warning on my own test: that was one fairly easy extraction, not a full corpus. It shows local isn't too weak — it doesn't prove local is always enough. JobHinge shelved local parsing for real accuracy reasons on real data. This doesn't overturn that; it says it's worth re-testing.
Their trial agreement states access is "for limited trial purposes only and without use of the API Service or Generated Content in production." Commercial use officially requires an NVIDIA AI Enterprise license.
This matters because JobHinge, Daily Soul Feed and Lottastar are live products, not experiments. I'd treat this as a real constraint on where we put it — not a technicality to wave off.
Four of the twelve endpoints I tried today were already dead — one retired ten days ago. In April, NVIDIA retired a model and its official replacement wasn't live yet, stranding users. Anything we build on this needs a fallback, or it breaks on a Tuesday for no reason.
NVIDIA says content isn't stored past your session "except in specific disclosed cases," and wouldn't say what those cases are when asked directly on their forum. Independent write-ups are blunter and warn against sending confidential data.
Practical rule: job listings, scripture, public business listings, image prompts — fine. Anything from your Freshservice work calls, anything with a real person's private details — never.
Ranked by what I'd actually do, not by what's technically possible.
| Project | The opportunity | Value |
|---|---|---|
| JobHinge | Meaning-based job matching. Right now the board matches on keywords. Free embeddings + reranking would let someone search "remote python work near Tampa" and get genuinely relevant jobs ranked properly. Your database already has the pgvector columns sitting unused. Both models tested clean and fast today. | BEST FIT New capability, not a swap |
| RCP Lead-Gen | Smarter company sorting. The corridor miner uses regex and rules — there's zero AI in it today. A free model could judge "is this actually a business without a website?" and de-duplicate companies far better than string matching. Nothing to lose, no current spend to protect. | BEST FIT Pure addition, $0 today |
| Daily Soul Feed | Free images worth testing. FLUX.1-dev produced a real image in 6 seconds at no cost — DSF currently pays OpenAI for every image. Worth a side-by-side, but your character-consistency work is tuned to the OpenAI model, so this is a test, not a switch. Script writing could also be A/B'd. | TEST FIRST Quality is the whole product |
| Aesop | Build it on NVIDIA from day one. The IT story channel hasn't shipped yet and is planned on paid OpenAI. Story remixing is exactly the kind of non-urgent text work the free tier handles well. Cheaper to start here than to migrate later. | EASY WIN Nothing to break |
| Talos & Nami | Less than I claimed — see section 04. Your boxes already beat the free cloud on extraction work, faster and just as accurately. Use the API for what you have no local answer to: images, reranking, embeddings, long documents, reading pictures. Not as a general upgrade. | NARROWER Corrected after testing |
| Scribe | Skip it. Technically the summary step could use a stronger free model — but Scribe handles your employer's work calls. Given the privacy caveat above, this is the one place I'd say no outright. | DON'T |
| Lottastar / Kepler | No AI in it — it's pure statistics. Nothing to gain. | N/A |
You spotted this, and it corrects something I told you. I said self-hosting NVIDIA models was impossible for you. It isn't — it's just not free.
Brev is not a cloud. It's a middleman that rents you GPU time from Amazon, Google and Lambda Labs behind one console. NVIDIA bought the company in July 2024.
Being in NVIDIA's Developer Program gives you a free license to run their models yourself on up to 16 GPUs for development and testing. That buys the two things the free API can't: no 40-per-minute limit, and your data never leaves the machine. Production use costs $4,500 per GPU per year — you're nowhere near that.
| What you'd run | Rough rate | 2 hrs/day | Left on 24/7 |
|---|---|---|---|
| A 70B model, privately | $1.10–2.00/hr | ~$66–119/mo | ~$800–1,400/mo |
| Image / video generation | ~$1.00/hr | ~$61/mo | ~$727/mo |
| Occasional batch jobs | $0.05–0.17/hr | ~$3–10/mo | trivial |
1. NVIDIA's published prices don't add up. Their table lists an A100 80GB at $6.21/hr — Lambda Labs sells the same card directly for $2.79. It also lists an H100 cheaper than Lambda's own price. That table looks stale. Get a live quote in the console before believing any number.
2. There is no idle auto-shutdown. Turn on auto-recharge, forget an instance is running, and it bills your card indefinitely. That's how people get hurt here.
3. Brev isn't the cheapest. Vast.ai and RunPod both beat it on identical hardware and bill by the second.
Not because you can't — because you have no problem that needs it. Renting only makes sense when all three are true at once: the model is too big for Talos and Nami, and you need privacy or no rate limit, and you'd use it in short bursts rather than around the clock. Today your paid stack covers the gaps and — per section 04 — your own boxes handle more than I credited them for.
The one thing that would change my answer: wanting to run something genuinely large against private data — anything touching your work calls, or a model tuned on your own JobHinge corpus. That's precisely what the free tier forbids and your hardware can't hold. Bookmark it for then, at a few dollars an hour, not as a standing subscription.
Short answer: read about four files, don't adopt the codebase.
NVIDIA/GenerativeAIExamples is real and actively maintained (4,142 stars, commits this week) — not abandoned demo-ware. But roughly 90% of its bulk is enterprise GPU-deployment scaffolding: fine-tuning notebooks, Kubernetes packaging, digital humans, warehouse robotics. All dead weight for you.
The parts that actually run with just your API key and no GPU:
| Path | What you'd get from it |
|---|---|
| RAG/examples/basic_rag/langchain/ | The exact wiring — model names, endpoint setup — to light up JobHinge's unused pgvector columns with real semantic search and reranking. The single most useful thing in the repo for you. |
| oss_tutorials/Nemotron_Parse_.../ | A working PDF-reading pipeline on the hosted API. Copy it whole if lead-gen or JobHinge ever needs to read documents. |
| RAG/examples/advanced_rag/structured_data_rag/ | Ask plain-English questions of a spreadsheet. Useful for ad-hoc reporting over your own exports. |
| community/5_mins_rag_no_gpu/ | Built explicitly for people without a GPU. Good first thing to run. |
Their reference build wants Docker, LangChain and a separate Milvus database. You don't need any of it. Your Postgres already has the pgvector columns. Roughly 50–150 lines calling the embed and rerank endpoints directly gets you there faster with nothing extra to maintain. Read their example for the pattern, then write our own.
Being straight with you rather than filling gaps: