One OpenAI-compatible API for today's open models: large models that see and reason, mid-size models for everyday work, and small models running your own adapters. Soon, your own machines can join the fleet.
We follow the open models people actually use and add them as they are released, so the list keeps moving. The names below are examples; your key comes with the current model ids and prices.
The largest open mixture-of-experts models, reading images and reasoning at length, for the hardest questions and agent work.
Fast, capable models for chat, tool calling, coding and documents: the default for most applications.
Small, inexpensive models, each with any of its LoRA adapters from Hugging Face. Name the adapter's repo as the model id; first use fetches it, every request after applies it instantly.
Coming next: embeddings for search and retrieval, and image generation Coming soon
Standard OpenAI SDKs work unchanged: set base_url
to https://api.solverfleet.com/v1 and use a model id that comes
with your key.
# Python, with the official OpenAI SDK from openai import OpenAI client = OpenAI( base_url="https://api.solverfleet.com/v1", api_key="YOUR_KEY", ) reply = client.chat.completions.create( model="MODEL_ID", messages=[{"role": "user", "content": "Summarise this contract."}], ) print(reply.choices[0].message.content)
JSON schema enforced while decoding; the reply is valid by construction.
"response_format": {"type": "json_schema", "json_schema": {
"name": "result", "schema": {"type": "object",
"properties": {"city": {"type": "string"}},
"required": ["city"]}}}
An unsupported schema is refused with a 400 up front, never silently ignored.
Choose how much a model thinks, and pay for depth only when it helps.
"reasoning_effort": "minimal" // minimal | low | medium | high
Or send "max_thinking_tokens": 512 for an exact cap.
OpenAI-format tool_calls, streaming or not.
Send tool results back as role: "tool" messages.
Images and whole PDFs through the same chat API, on models that see.
Every PDF page is rendered and passed in order, up to 32 pages.
detail: "low" takes a full-page scan from about 1,400
input tokens to about 200.
Shared context is cached, so agent loops answer in a fraction of the time.
On our serving machine, an image conversation's first turn takes 1.6 s to the first token and each turn after it 0.2 s.
Token counts on every response, and billing matches them.
The software that runs this fleet will run on yours. Join with the Macs, Linux and Windows machines you already own, and they work as one: for your own work first, and for this API when you choose.
Your apps ask for "chat" or "embeddings"; whichever machine can answer does.
Each machine gets the models and engine build its memory and chip can run, made ready in one act.
Each machine measures itself and keeps only the settings that make it faster.
A chat app, embeddings for your documents, image generation and agents that work while you are away.
Use the model files you already have, and serve your own fine-tunes beside the catalogue's.
Pair a phone by scanning a code. Each machine gets its own secure name, so phones and browsers connect without certificate warnings.
Models, conversations and keys stay on your network, and the fleet keeps working when the internet does not.
If you choose, your machines serve this API while they would sit idle, and are paid for every token. Your own work always comes first.
One installer for macOS, Linux and Windows. No developer tools needed.
Machines on your network find each other; machines elsewhere join with a code.
From your apps, your browser or your phone, the machine that can do the work does it.
Prompts and completions are processed in memory and not stored. Billing records keep a request id, timestamps and token counts, never content, and we keep the email address you write from. Nothing you send is used for training, sold or shared. Requests are served from our own hardware; nothing is subcontracted to third-party compute.
The API is in early access and provided as is, with no uptime guarantee yet. You are billed for the tokens you use at the prices that came with your key, and you may stop at any time. Do not use the service for anything unlawful or to harm others. Each open model remains under its own licence, which applies to what you do with its output.
For an API key, write to contact@solverfleet.com with a sentence about your use case. To join with your machines, write to the same address and we will tell you when it opens.