150,000,000 output tokens, free. Verify a mobile number and start generating — no credit card, no trial clock. After that, run your own open-weight models with unlimited tokens from ₹4,999 a month. No meter, nothing to ration.
Verify an email and a mobile number and you get 150,000,000 output tokens free — enough for weeks of real work, not a demo. Point Codex, Cline, Aider, Continue or anything OpenAI-compatible at our endpoint and start generating. No credit card, no trial clock.
Output tokens only — we never bill you for the context you send, because charging for input punishes exactly the long-context coding this model is good at. Runs at 60 token/sec on the free tier.
Share your invite link. Once they have generated 100,000 tokens — about one afternoon — you both get another 150,000,000. It pays out on real use, not on signups, so the number means something.
Alibaba’s newest open model, released August 2026, on our own hardware in India at a 128K context — and it reads images, not just text. An OpenAI-compatible endpoint and an API key — nothing to install, nothing to configure.
A verified mobile number is required — it is what stops one person collecting the grant a hundred times, and it is the only thing we ask for. Free tokens run at 60 token/sec; paid plans start at a 60 token/sec floor.
We ran the model against the work you would actually give it and checked the answers by running them — generated code executed against assertions, SQL against a real database, the concurrency answers under fifty threads. Here is what came back.
| Task | How we know |
|---|---|
| Web pages and UI | built this site’s demo page in 49 seconds |
| REST APIs, CRUD, auth middleware | idempotent payment handler survived 40 threads delivering the same webhook |
| SQL that is safe by default | parameterised a filter containing '; DROP TABLE users; -- |
| Concurrency-safe code | counter stayed exact across 50 threads × 2,000 increments |
| Fixing bugs from a stack trace | diagnosed an nginx 502 as a proxy read timeout from the log line alone |
| Writing tests for existing code | its own assertions passed against known-good code |
| Refactoring without breaking things | collapsed duplicated branches, behaviour identical on five cases |
| Messy real-world data | CSV with a BOM, quoted commas, escaped quotes and blank lines |
| Docs, READMEs, release notes | four required sections, in order, with a runnable sample and no invented URLs |
| Hindi and Indic languages | translated product copy into Devanagari, not English |
Told up front, because finding a limit yourself after signing up is a worse experience than being warned about it.
| Not this | Why |
|---|---|
| Whole-repository architecture changes | holding twenty files in mind at once is where a 30B is genuinely weaker |
| Long-form editorial to an exact word count | it drifts past a hard 300-word brief |
| Novel algorithm or protocol design | reach for a frontier model, and keep us for the other 80% of the work |
Think of it as a capable mid-level engineer: excellent on well-scoped
work, and reviewed like any colleague’s. It runs inside Claude Code,
Codex or Cline, so tests and diffs catch mistakes before they land.
The test suites are open — eval_model.py and
eval_realworld.py in our repo. Run them yourself.
Commit messages. Docstrings. Test scaffolding. Boilerplate, config, release notes, the fourth refactor of the same function. Most of what a coding agent emits all day is work a 30B model does correctly — and every one of those tokens comes out of the same budget you needed for the hard problem on Thursday.
On a $200-a-month plan the limit that bites is not the bill, it is the wall. You hit it on Thursday because Monday went on autocompletions. That is the expensive part — not the money, the moment you needed the good model and had none left.
A four-person team emitting 8M output tokens spends about that on the routine 80% alone, at Opus list prices. The same work here is ₹4,999 flat, or nothing at all while your free grant lasts.
We are not asking you to give up Claude. Keep it for architecture and the genuinely hard twenty percent. Point everything else here and stop rationing. Both tools, one editor, no wall.
Measured on our own endpoint, this is what “trivial” costs:
| Model | Compared with | Their price | Ours, flat | You save |
|---|---|---|---|---|
| Qwen3.8-27B | Scaleway, dedicated | $0 | $52 | |
| Llama 3.1 8B · Qwen 3.5 9B | Together, reserved dedicated | $0 | $52 | |
| DeepSeek-R1-Distill-Qwen-14B | Together, reserved dedicated | $0 | $58 | |
| DeepSeek-R1-Distill-Qwen-32B | Scaleway, dedicated | $0 | $115 | |
| DeepSeek-R1-Distill-Llama-70B | Fireworks, dedicated | $0 | $1,047 | |
| DeepSeek-R1 671B | Fireworks, 8-GPU cluster | $0 | $7,853 |
The moment a prompt works and you want it across the whole corpus, a metered API bills you more for succeeding. Add an agent loop or a reasoning model and the curve gets steeper. You end up rate-limiting yourself before the provider does.
No requests-per-minute cap, no tokens-per-day quota, no tier to upgrade and no 429 at 3am mid-batch. The card is yours, so you run it flat out for as long as you want.
Thinking traces cost two to three times the tokens you actually see. On a metered plan you pay for every one. Here the number doesn't appear on your invoice, so you can leave reasoning on.
Standard /v1/chat/completions served by vLLM. Point your
existing OpenAI client at a new base_url, keep every other line
of code, and roll back just as easily.
# same client, same code — one line changes
client = OpenAI(
base_url = "https://<you>.optracloud.com/v1",
api_key = "sk-…",
)
resp = client.chat.completions.create(
model = "deepseek-r1-distill-qwen-14b",
messages = [{"role": "user", "content": "…"}],
)Shared plans are capped at 128K so one long session cannot crowd out everyone else. On a dedicated card the whole memory is yours — 256K, 512K, or a full 1M tokens, depending on the tier. Pick the model you want to run and the context you need; we size the hardware and you never think about it again.
| What you run | Context window | Your speed, alone | Per month |
|---|---|---|---|
| Up to 14B | 256K | ~60 token/sec | ₹24,499per month |
| Up to 32B — incl. Qwen3.8-27B | 256K | ~60 token/sec | ₹43,499per month |
| Up to 70B at 4-bit | 128K | ~75 token/sec | ₹53,499per month |
| 120B MoE · 70B dense | 512K | ~120 token/sec | ₹95,499per month |
| 120B MoE, low-latency tier | 512K | ~255 token/sec | ₹206,499per month |
| 235B MoE and up, multi-card | 1M | ~365 token/sec | ₹265,999per month |
Speeds are measured on our own hardware with vLLM and 4-bit weights, for a single user with the card to themselves. Context windows are the ceiling for the largest model in each tier — run something smaller and there is room to spare. Every tier is yours alone: no neighbours, no queue.
Every instance is a dedicated GPU allocated to one customer, billed flat with no token metering. The middle column is the cheapest published dedicated plan for the same model anywhere else. Pick one and tell us the workload — we will be in touch within one business day to agree scope and timing.
| Model | Context | Throughput | Same thing elsewhere | With OptraCloud |
|---|
Nothing selected yet. Choose any model above to start a request — we schedule hardware against confirmed demand, so telling us early is what gets a model built.
A metered API is cheaper for light usage. Move the slider to find your crossover.
| OptraCloud | Serverless token API | Self-hosting | |
|---|---|---|---|
| Rate limits / quotas | None | RPM + TPD tiers | None |
| Hardware shared with others | No | Yes | No |
| Predictable monthly cost | Yes | No — per token | Yes |
| Single-tenant hardware | Yes | No | Yes |
| Upfront capital | None | None | $5k – $280k |
| Engineer needed to run it | Ours | None | Yours |
| OpenAI-compatible API | Yes | Yes | You build it |
| Swap models without new hardware | Yes | Yes | No |
| Cheaper at low volume | No | Yes | No |
| Someone to call at 2am | Yes | Ticket queue | You |
OptraCloud is the AI infrastructure arm of Globussoft Technologies, incorporated 2012. We run the facility, announce our own address space, and hold the certifications your security review will ask for.
Certificates are held by Globussoft Technologies Pvt. Ltd. — ISO/IEC 27001:2022 for information security, cyber security and privacy protection (cert IC-58930/1225, valid to December 2028) and ISO/IEC 27018:2019 for personally identifiable information in public cloud (GCS/ITST/8532). We send full certificates and scope statements on request, before you commit to anything. 300+ engineers, serving over two million users across 50 countries since 2012.
Billed in dollars or rupees, whichever suits your accounting. Indian customers get GST invoices on every bill with no forex markup and no card-network conversion fee.
No reservation, no minimum term, no cancellation fee. Annual prepay earns a discount if you want one, but nothing on this site requires a commitment to get the price you see.
Flat plans do not meter tokens, render seconds or requests. Hourly billing stops the second you destroy the instance. Overages, where they exist, are published on this page.
| Is my data used for training? | No. Nothing you send is logged, retained beyond the request, or used to train anything. On dedicated plans the weights and the machine are yours for the term. |
| What happens if I outgrow the plan? | Move up a tier or add a card. Same endpoint, same API keys, no migration and no re-onboarding. |
| Can I bring my own model? | Yes. Any open-weight model that fits the tier you pick, plus your own fine-tunes and LoRAs. Ask and we load it. |
| Where is the hardware? | One facility we own and operate, with dual-carrier transit and our own ASN. If your contract requires a specific jurisdiction, we can commit to it in writing. |
| What is the SLA? | Availability targets go in the contract. Ask during evaluation and we will be specific about what we will sign. |
| Do you offer a trial? | Shared plans start at $52 a month and cancel any time, which is a cheaper way to test quality than any trial we could design. Hourly GPUs start at $0.39. |
We come back with the hardware it needs, the monthly cost, and your provisioning date — usually within one business day.
No card required. We reply within one business day — [email protected] · +91 74063 17771
We have your request and a named engineer will be in touch within one business day to talk through what you need.
—