open ai models.
distributed gpu pool.
One OpenAI-compatible API. Open-source models. Served by a verified pool of GPUs across Europe — from gaming rigs to datacenters. Builders pay less. Hardware owners get paid for every token their machines serve.
How ElephantPool works
Developers bring requests. Hardware owners bring GPUs. Our scheduler, verifier and payment rails sit in the middle — and take 20%, only when tokens flow.
Builders call the API
Swap your base URL, keep your code: an OpenAI-compatible endpoint for Llama, Qwen, DeepSeek, gpt-oss. Pay per token, prepaid credits, no subscription.
The pool serves it
The scheduler routes each request to the best verified machine — by speed, price, reputation and (if you choose) EU-only residency. Every job produces a cryptographic receipt.
Owners get paid
80% of every token's revenue
The split is public and applies to every job on the network.
Models & pricing
The latest open models, transparent per-token pricing, in euros. The EU-Sovereign tier routes only to machines located in the EU, checked by us rather than declared by them — and still undercuts every sovereign cloud.
| Model | Standard € / M tokens in · out |
EU-Sovereign € / M tokens in · out |
Elsewhere cheapest US · EU sovereign |
|---|---|---|---|
| Llama 3.1 8B 8B dense · 128K |
€0.03 · €0.06 | €0.05 · €0.10 | US $0.02 · $0.04 DeepInfraEU — |
| Qwen3.8 27B 27B dense · 262K · Apache-2.0 · new |
€0.25 · €1.50 | €0.35 · €2.00 | US $0.40 · $3.00 ChutesEU — |
| Gemma 4 26B-A4B 26B MoE · 256K · Apache-2.0 |
€0.06 · €0.25 | €0.10 · €0.35 | US $0.07 · $0.34 DeepInfraEU — |
| Llama 3.3 70B 70B dense · 128K |
€0.30 · €0.50 | €0.35 · €0.55 | US $0.10 · $0.32 DeepInfraEU €0.65 · €0.65 IONOS |
| gpt-oss-120b 117B MoE · 131K · Apache-2.0 |
€0.06 · €0.30 | €0.10 · €0.45 | US $0.03 · $0.17 CoreWeaveEU €0.08 · €0.40 OVHcloud |
| DeepSeek V4-Flash 304B MoE · 1M · MIT |
€0.15 · €0.35 | €0.25 · €0.55 | US $0.08 · $0.18 DeepInfraEU €0.40 · €0.80 Scaleway |
| MiniMax M3 427B MoE · 1M · community license |
€0.25 · €1.00 | €0.35 · €1.30 | US $0.28 · $1.10 DeepInfraEU — |
| GLM-5.2 753B MoE · 1M · MIT |
€0.45 · €1.40 | €0.60 · €1.80 | US $0.48 · $1.49 NovitaEU €1.80 · €5.50 Scaleway |
| Qwen3.5 397B-A17B 397B MoE · 262K · Apache-2.0 |
€0.40 · €2.50 | €0.50 · €3.00 | US $0.30 · $1.93 DigitalOceanEU €0.60 · €3.60 Scaleway/OVH |
Target launch pricing. Batch tier: −50%. Competitor prices from public price lists, August 2026. Free starter credits on small models for early users.
All catalog models are open-weight (Apache-2.0, MIT or Llama/community licenses — linked per model in the API). Network-scale giants (Qwen3.8 2.4T, Kimi K3, DeepSeek V4-Pro) require multi-node serving and are on the roadmap, not the launch catalog.
Drop-in replacement
OpenAI-compatible /v1/chat/completions
GDPR by construction
Prompts never touch a disk, are never used for training, and the EU-Sovereign tier never leaves machines we have placed in the EU. DPA available.
Smart routing, smart savings
Pick a model — or pick auto
Your GPU has a day job now
Install mahout
RTX 3090 / 3080
RTX 4090 / 5090
RX 7900 XT / XTX
A6000 · A100 · H100
Safe by design
mahout is open source. It runs only signed engines in sandboxed containers — read-only weights, zero network egress, never any client code. No inbound ports on your router.
Your machine, your rules
Set availability windows, power limits, disk quota and a price floor. Auto-pause when you game or work. One-click stop, always.
Paid for work, monthly
80% of every token your machine serves, tracked by cryptographic receipts you can verify yourself. Monthly SEPA payout from €50, processed by Stripe. Guaranteed starter work in your first week.
What could your hardware earn?
Estimates for continuously batched serving at our launch price book, with 80% of token revenue paid to you. You are paid for work actually routed to your machine — this is an estimate, not a promise.
Assumptions used:
- Launch price book, continuous batching
- Your share: 80% of token revenue
- Power draw under load per rig shown model-by-model
- Utilization depends on network demand in your region
⚠️ An estimate, not a promise. You are paid for tokens your machine actually serves — volume follows network demand. An estimate at the utilisation you chose, not a promise of demand. Measured earnings — from a benchmark on your own card — live in your dashboard.
Don't trust us. Verify us.
Verified compute
Every job produces an activation fingerprint (the TOPLOC scheme — 258 bytes per 32 tokens). A random sample of all work is replayed on trusted hardware. Machines that fake work don't get paid — and get banned.
Public receipts
Signed job receipts are batched into a Merkle tree and anchored on a public blockchain every 10 minutes. Any host or client can independently prove what was served, when, and what was owed.
Open source, open docs
The mahout agent is open source and auditable. Our architecture, economics and verification design are published: read the whitepaper
European, by choice
Operated from Belgium by De Buck Technologies. EU-Sovereign tier with enforced EU-only routing, zero retention, and a standard DPA — at roughly half the price of sovereign clouds.
One public split: 80 / 20
80% of token revenue to the machine that did the work, 20% to the network. No hidden spreads, no house hardware jumping the queue on equal terms.
Built to compound
Long term, idle pool capacity trains open models collectively — the network that serves AI also improves it, in the open. That roadmap is public too.
Frequently Asked
Is this safe for my PC?
mahout runs only pinned, checksum-verified artifacts — a specific engine release and specific model files, each verified while downloading and again at every load. It runs as your user, needs no root, and opens no inbound port. It is open source, so you do not have to take our word for it. Auto-pause keeps it out of your way while you are using the machine. Being honest about the current limit: the engine runs as an unprivileged child process bound to localhost, not yet inside a syscall sandbox.
How and when do I get paid?
80% of the revenue of every token your machine serves. Monthly SEPA payout from €50
What if my internet or PC is slow?
mahout benchmarks your machine and the network only routes work it can genuinely handle — smaller models on smaller rigs. Slow uplinks serve batch jobs instead of live chat. Everything is measured, nothing is promised that your hardware can't do.
Is AMD really fully supported?
Yes — same features, same tiers, same pay as NVIDIA. Today every card runs through one Vulkan build of llama.cpp, which serves AMD, NVIDIA and Intel alike; native ROCm and CUDA builds are a performance step we are working on, not a prerequisite. Our own fleet includes RX 6900 XT and RX 7900 XTX cards, so the AMD path is exercised daily.
Is my prompt data private?
Prompts and completions live in RAM only, are never logged, and are never used for training
Why open-source models only?
Open weights are what make a distributed pool possible — anyone can verify what's running, and nobody can take the models away. They're also what makes it cheap: no license margin, just hardware and electricity.
Read before you trust us
What runs on a host machine, what a GPU actually earns, where prompts go, and how a request finds a card. Written plainly, including the parts that are still work in progress.
a GPU died mid-sentence. the sentence finished.
Consumer hardware disappears without notice — someone sits down at their PC, a breaker trips, Wi-Fi drops. Our streams now survive that: another machine picks the answer up mid-word, and the dead segment is billed to nobody.
Read the article →one model, every card: how a 27b runs on 8 gb
Qwen3.8 27B ships in five builds, from Q5_K_M on a 24 GB card down to IQ2_XXS on 8 GB. Why we split models from builds, what quantization actually costs, and how a quantized KV cache doubled our context overnight.
Read the article →we stopped projecting your earnings and started measuring them
Most GPU marketplaces advertise what a card could earn. Our dashboard shows what yours measured, and refuses to show a number until it has one. Here is why that choice cost us a nicer-looking page and was still right.
Read the article →Knowledge base
Short answers to the questions people actually ask.
- What actually runs on your machine
- Which model can my GPU serve?
- What your GPU actually earns
- Security and privacy: what we do, and what we don't
Ready to dive in?
Builders: free starter credits on small models. Hosts: install mahout and your GPU earns on every token it serves.
Get API Access Connect your GPUOne account for both sides.