By day your machines work for your people; by night they become a team of AI agents on a 70B model running inside your network. Hand them the task when you head home and it's done by morning. No GPUs to buy, no cloud bill, and your data never leaves the building.
Not a chatbot waiting for questions: agents that read, decide, use tools and hand you finished work when you arrive.
Reads hundreds of contracts, flags risky clauses, deadlines and penalties, and leaves you a report for each one.
Sorts, prioritizes and drafts the first reply to everything that came in during the day. Your team starts the morning with an orderly inbox.
Walks your repository, writes tests and does code review with a coding model. Leaves it ready for your team to review.
Invoices, scans and drone photos turned into structured data and a report ready by morning.
They speak the OpenAI format with tool calling, so they work with open-source agents. Every model call leaves an auditable receipt and nothing leaves your network.
You control who computes, and every job is audited. Not a single byte goes to a third party.
The swarm runs on your local network. Prompts and answers never go to the internet or through a third party. Air-gap deployable.
Every job leaves a receipt with SHA-256 fingerprints (not the text): prove what was computed and who served it without revealing the data.
In a hurry? Co-located power on a dedicated GPU (~30–80 tok/s). No rush? Batch swarm, far cheaper. Same endpoint.
We don't sell speed: we sell cheap access to a big brain for batch work that isn't in a hurry. These are the cases that pay, with the model that serves them.
| Case | Model | Why it fits |
|---|---|---|
| Synthetic data train other AIs | Llama-3.3-70B | Inherently batch; the model's quality is the data's value. The sharpest wedge. |
| Bulk summaries legal, minutes, tickets | Qwen2.5-72B | Thousands of documents overnight; nobody watches the screen while it runs. |
| Contract analysis risks and dates | DeepSeek-R1-70B | Step-by-step reasoning a lawyer can audit. |
| Documents and images invoices, drones | Qwen2.5-VL-72B | Structures data from PDFs and photos in batches: agriculture, inspection, insurance. |
| Classification / PII moderation, dedupe | Mixtral 8x7B | Millions of items, cheap thanks to MoE (few active parameters). |
| Translation / catalogs localization | Qwen3-30B-A3B | Massive volume at minimal compute. Ideal for cheap on-prem. |
| RAG over your archive internal search | BGE-M3 | Overnight indexing of entire libraries, without the data leaving your network. |
Everything runs inside your network and leaves a hash-auditable receipt. You control who computes.
Change one line of base_url and you're running. Same OpenAI SDK, open models, no GPUs to buy or maintain.
Whoever rents out their GPU loads only part of the model. Nobody needs a huge machine.
Your text runs through the layers across several nodes and comes back as an answer. Like a torrent, but for compute.
With no hardware to maintain, the cost collapses. You pay for what you use.
An open 32B–70B model approaches the quality of expensive APIs, but runs on your hardware and costs cents. Power isn't a number to brag about: it's what lets you build without asking anyone's permission.
With function/tool calling (OpenAI format) the model reads, edits and tests code in a loop. Leave it the task when you go home; it's done by morning.
Classify, summarize or extract data from your whole archive overnight on the batch lane. The volume that costs a fortune on a closed API costs cents here.
Ask natural-language questions against your manuals, contracts or case files. The data never leaves your network: the model goes to the data, not the other way round.
Tag tickets, sort emails, read invoices. Simple tasks but by the millions: exactly where price per token decides whether the project is viable.
Pick the model with model on every call. Chat, code, vision, reasoning or cheap volume: whichever fits your case, all on the same endpoint.
| Model | Type | Good for |
|---|---|---|
| Llama-3.3-70B | chat | Strong reasoning, legal/financial analysis, synthetic data |
| Qwen2.5-72B | chat | High-quality all-rounder: complex summaries, Q&A, writing |
| Qwen2.5-32B | chat | The sweet spot: structured extraction, fine summaries, cheap |
| Qwen2.5-Coder-32B | code | Repo analysis, tests, code review, tool-calling agents |
| DeepSeek-R1-Distill-70B | reasoning | Step-by-step analysis and decisions that must be justified |
| Qwen2.5-VL-72B | vision | PDFs, invoices, scans, drone photos → structured data |
| Mixtral 8x7B | MoE | Bulk classification and translation: millions of items, dirt cheap |
| Qwen3-30B-A3B | MoE | Massive volume (tagging, localization) at minimal compute |
| Gemma-2-27B | chat | Writing and summaries with a good quality/cost ratio |
| Llama-3.1-8B | fast | Simple tasks at high speed and minimal cost |
| BGE-M3 | embeddings | Index your documents for search / RAG |
Need another one? The catalog grows without touching code. Just ask.
It's the OpenAI API. If you already use their SDK, you change two lines: the base_url and your key.
# pip install openai from openai import OpenAI client = OpenAI( base_url="https://enjambreia.com/v1", api_key="sk-enjambre-...", ) r = client.chat.completions.create( model="qwen2.5-72b", messages=[{"role":"user","content":"Hello"}], ) print(r.choices[0].message.content)
curl https://enjambreia.com/v1/chat/completions \ -H "Authorization: Bearer sk-enjambre-..." \ -H "Content-Type: application/json" \ -d '{ "model": "qwen2.5-coder-32b", "messages": [ {"role":"user","content":"Write tests"} ] }'
Same response format as OpenAI: choices, streaming and tool-calling included. Your current libraries work unchanged.
Same job, price per 1M input tokens. The longer the bar, the more expensive.
Reference figures. The closed API uses public high-end pricing; the own GPU assumes a card saturated 24 h a day (if it sits idle, it costs more). Adjust to your case.
We don't manufacture hardware or build warehouses full of GPUs. We turn compute that already exists —your idle machines or a rented GPU— into private AI. No capex, no construction, no waiting years.
Zero investment in GPUs or installation. Pay for what you use; switch on power only when you need it.
Runs on your existing or rented hardware. Nothing to manufacture, nothing to install in your backyard.
Change one line of base_url and you're running. No waiting for infrastructure to be deployed.
Creating an account is free and there's no subscription: you pay only for what you use, from cents per million tokens. If you use your own GPU (Personal Node Plan) you pay only a 20% fee — the closest thing to free AI, no tricks.
enjambreia spreads compute across GPUs that already exist, so it doesn't pay for data centers: open 70B models from $0.10 per million input tokens, up to 25× less than a high-end closed API.
Yes. It's compatible with the OpenAI API: change the base_url and your key, and your code, SDK and tools (like OpenCode) keep working, with streaming and tool-calling.
Yes: install the node client and lend your GPU (half by default). While it's on and approved, you use the swarm's models at no cost, including the ones split across several cards. Cloud models are paid separately.
Every job is signed and audited, and companies and private groups only use their own machines: your information never goes through outside nodes.
Llama 3.3 70B, Qwen2.5 72B, Qwen2.5 Coder, DeepSeek R1, vision models and whatever the nodes announce. The current list is at /v1/models.
Create your account, top up your balance and pay only for what you use.
Sign up