- Attention size
- 1M tokensshared by all requests in flight on Qwen3.6-35B-A3B on ARK Cloud
- API dialects
- OpenAI + Anthropicone key, same models, change the base URL
- Modalities
- Text, vision, audiochat, reasoning, tools, embeddings, speech to text
- Where it runs
- EU or on-premARK Cloud in the EU, ARK Core on your GPUs
Keep your code. Change two lines.
ARK speaks the APIs your stack already uses. Point the OpenAI or Anthropic SDK at ARK, pick an ARK model, and the rest of your application stays as it is.
- OpenAI: chat completions, embeddings, audio transcriptions, model listing
- Anthropic: messages and token counting
- Streaming with usage, tool calling, reasoning, image input
- Stateful sessions: send only the new turn, ARK keeps the conversation
client = OpenAI(
- base_url="https://api.openai.com/v1",
+ base_url="https://api.ark-labs.cloud/api/v1",
api_key=os.environ["ARK_API_KEY"],
)
response = client.chat.completions.create(
- model="gpt-4o-mini",
+ model="arklabsai/Qwen3.5-122B-A10B",
messages=messages,
tools=tools,
)Open-weight models, ready to call.
Reasoning and vision models up to 122B parameters, a Polish-first LLM, multilingual embeddings and speech recognition. Prices per million tokens.
Start in our cloud. Move it in-house when you are ready.
The API, the models and the behaviour are the same wherever ARK runs, so a prototype built on ARK Cloud becomes a self-hosted service without a rewrite.
ARK Cloud
Pay-as-you-go API with prepaid credits. Sign in with Google, take your key, send requests.
- Six models, OpenAI- and Anthropic-compatible
- Per-token pricing, no subscription
- Hosted and operated in the European Union
ARK Core
The ARK platform on your own infrastructure, on-premises or in the GPU cloud you choose. A model can span several GPUs and machines.
- The same API and models as ARK Cloud
- Runs on NVIDIA GPUs, from RTX cards to the data centre
- Your network, your logs, your keys
ARK Tailored
A deployment designed with our engineers: models, hardware sizing, integration and support.
- Capacity planned for your traffic
- Custom models and integrations
- Direct line to the team that builds ARK
Every GPU you have, working as one.
Bring machines with whatever NVIDIA GPUs they carry. ARK joins them over ordinary Ethernet into one deployment and splits each model across them, so cards of different generations and sizes serve the same model together.
- One model, three machines, two GPU generations: Qwen3.5-122B-A10B runs in production on four RTX 5090 in two machines and an RTX 4060 Ti in a third.
- Mix generations and sizes, from an RTX 3060 to data-centre GPUs.
- Grow by adding a machine: its servers join the running deployment and their models appear in the API, with no restart.
Throughput that grows with your traffic.
Measured on ARK Cloud through the public API with real summarisation traffic. Qwen3.5-122B-A10B on four RTX 5090 GPUs kept the median time to first token under 0.6 seconds up to 24 concurrent requests, and processed 5,726 tokens per second at 60.
- Qwen3.5-122B-A10B
- 5,726 tok/sinput + output at 60 concurrent requests
- Qwen3.6-35B-A3B
- 7,448 tok/sat 48 concurrent, on four RTX 4070 Ti SUPER
- First token
- 0.27 smedian for Qwen3.6-35B-A3B, one request at a time
- Reliability
- 730 / 730Qwen3.5-122B-A10B requests completed across the run
Throughput on ARK Cloud
Total tokens per second (input + output) as concurrent requests grow. Summarisation traffic, about ten input tokens per output token, October 2026.
Everything a production app asks of an inference API.
Reasoning
Turn thinking on with reasoning_effort or Anthropic thinking; reasoning tokens are reported apart.
Tool calling
OpenAI tools and Anthropic tool_use, streamed or not, with forced tool choice.
Vision
Send images with the prompt to the Qwen models, as URLs or base64 data.
Stateful sessions
Name a session and send only the new turn. ARK keeps the conversation; you pay for new tokens.
Embeddings
Multilingual dense vectors from bge-m3 for search, RAG and clustering.
Speech to text
Whisper large-v3-turbo transcription, with subtitles and timestamps.
Streaming with usage
Token-by-token SSE, a final usage chunk, and optional running counts on every chunk.
Anthropic Messages
Use Claude-style clients and SDKs against open models, including count_tokens.
Your prompts stay inside the perimeter you choose.
ARK Cloud runs in the European Union on hardware we operate ourselves. When your rules say data must stay on your premises, ARK Core and ARK Tailored put the same platform inside your own network.
- No training on your data. ARK serves models; it does not train them.
- Logs without content. Request logs keep metadata: prompt fields are redacted and generated text is not logged.
- Known models, published licences. Every model is an open-weight model you can inspect, with its licence on our model page.
- Your own perimeter when you need it. Run the whole platform on your infrastructure with ARK Core or ARK Tailored.
Agents, SDKs and apps connect in minutes.
Notes, research and news.
ARK Core under real traffic
Throughput, time to first token and reliability for every ARK Cloud chat model, from one request to sixty.
From prompt caching to ARK 2.1
The platform's big steps, newest first: prefill batching, paged and quantized attention, fused kernels, decode batching and more.
What made ARK fast, measured
Inverted processing loops, decode batching, fused kernels, paged and quantized attention, prefill batching, prompt caching: what each one improved, with the numbers.
Your first request is two minutes away.
Sign in with Google, copy your key, and call any ARK model with the SDK you already use.