< 100 ms TTFT
Time to first token, the metric your users actually feel, not brochure tokens-per-second. We lower it with prefix caching and chunked prefill in the engine.
token factory · llm api
Change one line. Your inference, in Europe.
Point the OpenAI or Anthropic SDK at GPUFlow and your code stays the same: 15 open source models served on NVIDIA B200 GPUs, in Spain, with balance in euros.
getting started
From signup to your first token in under 30 seconds.
- 01
Pick your balance
Start free with €5 when you create your account, or choose a bundle (10 / 50 / 200 / 1000 €) with bonus. Credit never expires.
- 02
Confirm payment
On the free trial you just confirm your account. For a bundle, you pay in Stripe's secure checkout, in euros.
- 03
Get your API key
It arrives by email along with access to the admin panel. Point your SDK at GPUFlow and run your first curl.
compatibility
You rewrite nothing. You change one line.
Token Factory speaks the OpenAI protocol (/v1/chat/completions) and the Anthropic one (/v1/messages). If you already use either SDK, you only change the endpoint.
- /v1/chat/completions and /v1/messages endpoints
- Streaming, function calling and configurable reasoning (reasoning.effort)
- Same request and response formats, no translation layer
endpoints
OpenAI
POST /v1/chat/completionsAnthropic
POST /v1/messagesbase url · https://api.gpuflow.ai
catalogue
15 open source models, deployed and ready.
DeepSeek, GLM, Qwen and Gemma, served on our own infrastructure. Price per million tokens, input/output split. Need another one and it fits on the GPU? We deploy it.
| model | capabilities | input · €/Mtok | output · €/Mtok | details |
|---|---|---|---|---|
DeepSeek V4 Pro 1.6T MoE on 8× B200. Full-power reasoning. ChatCodingReasoningTools | ChatCodingReasoningTools | 0.80 € | 2.60 € | details |
GLM 5.2 NVFP4 743B NVFP4 on 8× B200. Generalist and coding with tool calling. ChatCodingReasoningToolsLong context | ChatCodingReasoningToolsLong context | 1.00 € | 3.00 € | details |
DeepSeek V4 Flash 1.6T MoE on 2× B200. High throughput, 524K context. ChatReasoningTools | ChatReasoningTools | 0.14 € | 0.28 € | details |
Qwen3.6 27B 27B dense. Strong at coding and agents, with tool calling. ChatReasoningToolsOCR | ChatReasoningToolsOCR | 0.30 € | 2.40 € | details |
Qwen3 14B 14B FP8. Reasoning with a configurable thinking budget. ChatReasoningTools | ChatReasoningTools | 0.20 € | 1.00 € | details |
Qwen3.5 9B 9B ultra-light. Fast and cost-effective, with tool calling. ChatReasoningTools | ChatReasoningTools | 0.10 € | 0.15 € | details |
Gemma 4 26B A4B Multimodal MoE, ~3.8B active. Vision, reasoning and tools. ReasoningTools | ReasoningTools | 0.12 € | 0.30 € | details |
Nemotron Nano 12B VL Vision-language model for OCR and document understanding. 128K context, up to 4 images per prompt. image-text-to-textOCR | image-text-to-textOCR | 0.06 € | 0.24 € | details |
Qwen3-VL 32B OCR Vision-language model for OCR and document understanding. 64K context, up to 4 images per prompt. image-text-to-textOCR | image-text-to-textOCR | 0.10 € | 0.35 € | details |
Nemotron 3.5 Lightning 30B MoE + Mamba-2 hybrid, ~3B active. Reasoning and tool calling, 128K context. ReasoningTools | ReasoningTools | 0.12 € | 0.30 € | details |
performance
Why it's fast.
Time to first token is what your user feels. We look after it at the hardware level.
Dedicated NVIDIA B200
Latest-gen accelerators on bare metal. No shared tenant stealing your latency at peak time.
Internal RDMA network
Low latency between nodes. Combined with a Sandbox, calls never hit the public internet.
see it in action
One call, streaming response.
Compatible with the OpenAI and Anthropic SDKs: change the base URL, paste your key and you're calling 15 open source models served on B200 GPUs.
● live
from openai import OpenAI
client = OpenAI(base_url="https://api.gpuflow.ai/v1",
api_key=os.environ["GPUFLOW_API_KEY"])
stream = client.chat.completions.create(
model="qwen-3.5-9b",
messages=[{"role": "user", "content": "What is GPUFlow? Answer in one sentence."}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")edit the prompt and hit Try it
response
Live demo with usage limits · small-model output. For production, open your account.
pricing
Prepaid in euros. No surprises.
Load credit, consume by token, see spend in the dashboard. No fees, no minimums, no currency conversion.
- Credit from €10 · tiers with a welcome bonus
- Deductible Spanish invoice · valid for intra-EU reverse charge
- The same credit works for Token Factory, Sandbox and dedicated GPU
frequently asked
What developers ask before they start.
Do I have to rewrite my code?
No. Token Factory speaks the OpenAI protocol (/v1/chat/completions) and the Anthropic one (/v1/messages). Point the SDK you already use at https://api.gpuflow.ai, paste your key and the rest of your code stays the same.
Which models can I call?
All 15 in the catalogue (DeepSeek, GLM, Qwen, Gemma, NVIDIA, OpenAI, Mistral and Fish Audio), with a single API key. Need another open source model that fits the GPU? We'll deploy it.
Are there usage limits?
You consume against your prepaid balance: pay per token up to whatever you've topped up, no quotas or minimums. If you need sustained high volume, reach out and we'll raise your account limits.
Do you train on my data?
No. Your prompts and responses are processed to serve your request and are not used to train models. The details are in the privacy policy.
Where is my data processed?
In our datacenter in Spain, on dedicated NVIDIA B200 GPUs. Your data never leaves Europe.
Do you offer an SLA?
We publish availability on the status page. For a contractual SLA (enterprise accounts), reach out and we'll set one up.
€5 free for your first curl.
Create the account and call any model in the catalogue. No card.