llama-3.3-70b-self-hosted is now serving on our EEA GPU.
EU-hosted inference. Paid in BTC. No KYC. No subscription. Cancel by not topping up — there's nothing to unsubscribe from.
+30% bonus locked through Wed 20 May · 100% refundable in fiat (you cover exchange-rate delta) · Credits never expire.
Built for founders and engineers who can't afford to have their stack subpoenaed.
We don't compete on the cheap end. If your monthly spend is under $100, Together, Hyperbolic, or DeepInfra will serve you better and faster. Our customer is the founding engineer at a stealth startup, the AI consultant under NDA, the team running production inference where the billing trail is its own threat model.
On May 13 2026, Anthropic raised Max-tier token caps and announced a $200/mo Agent SDK credit (effective 15 June) — because users like our operator were burning through the prior limits: 40+ million tokens in 17 of the past 21 days. That isn't abuse; that's what production inference looks like in 2026. The credit closes the raw-throughput gap for Max customers — but the structural friction remains. No US card on file. No prompts in a US discovery surface. No passport handed to a reseller. If any of that describes your situation, you're our customer.
We target production-grade developers on large and x-large projects: teams who have hit the ceiling on Anthropic, Cursor, Windsurf, and Cline, and who need an independent path on jurisdiction, payment rails, and data retention — not just on tokens-per-minute. We're the viable option for the workload they can't (or won't) run through the default stack — privacy, jurisdiction, and billing trail first; smart-route cost savings second.
Pay in seconds with crypto. Cancel by not topping up. We never read your prompts. We never charge a card.
FIDO2 required at sign-in · YubiKey ships free · why
100% refundable in fiat until 70B EU model goes live · YubiKey included free
Built for teams at 40M+ tokens/month. Reserve capacity and lock in a lower per-token rate.
Direct comparison against routing every token to Sonnet 4.6 at retail.
| Daily workload | All-Sonnet 4.6 | llmdeal Starter | llmdeal Pro | Pro savings |
|---|---|---|---|---|
| 6M tok/day · light agent | $1,260/mo | $144/mo | $540/mo | $720/mo · 57% |
| 30M tok/day · steady prod | $6,300/mo | $720/mo | $2,700/mo | $3,600/mo · 57% |
| 100M tok/day · heavy agent fleet | $21,000/mo | $2,400/mo | $9,000/mo | $12,000/mo · 57% |
| Same Pro on Production commit (−20%) | — | — | + $499 base | $1.60/$4.00 net per 1M |
Assumes avg 800 input / 400 output tokens per request. Pro routing (~50% Qwen-Coder, 25% Llama-3.3-70B, 15% DeepSeek-V3.2, 10% Codestral/Qwen3-235B/GLM-5) bills at a flat $2/$5 per 1M weighted average — we absorb model-to-model cost variance internally. Starter is single-model (our self-hosted Qwen-Coder-32B), no routing overhead.
Pro per-bucket on 6M tok/day workload (4M input + 2M output):
50% to Qwen-Coder-32B (EU GPU): 2.0M in + 1.0M out → $9.00/day
25% to llama-3.3-70b-self-hosted: 1.0M in + 0.5M out → $4.50/day
15% to DeepSeek-V3.2 (reasoning): 0.6M in + 0.3M out → $2.70/day
10% to Codestral/Qwen3-235B (heavy): 0.4M in + 0.2M out → $1.80/day
TOTAL: $18.00/day · ~$540/monthPro routes exclusively across our self-hosted + open-weight stack (Llama, DeepSeek, Mistral, Qwen, GLM). Median savings versus a single frontier provider: ~55–65% vs Sonnet, up to 80%+ vs Opus-tier depending on workload mix.
Early builders who claimed a founder seat and pushed the product into shape before launch.
Software company shipping developer tools and AI workflow products — Snitch (security auditing), Jeremy (AI context layer), Scribe (voice-to-text), and a deep catalogue of focused developer and creative apps.
khuur.dev — precision software for developers and technical teams →
Real objections, straight answers — no sales spin.
The in-flight request completes — we absorb the overrun. Every subsequent request returns a 402 with an explicit "out of credits" body. Top up; service resumes immediately. No silent throttling, no surprise invoices.
Not today. Smart routing works because we hold the upstream contracts — that's what lets us route to the cheapest qualified model per request. BYO-key support is on the Sovereign tier roadmap, but it undercuts the routing margin, so it will be priced to reflect that.
A small open-source classifier (RouteLLM-style) scores each prompt on complexity, latency-sensitivity, code vs prose, and reasoning depth. Easy → Qwen-Coder-32B (our EU GPU). Fast workhorse → llama-3.3-70b-self-hosted (our EU GPU). Reasoning → DeepSeek V3.2 or Qwen3-Next 80B Thinking. Code-heavy → Codestral. Hardest queries → Qwen3 235B or GLM-5. Per-request telemetry shows exactly which model fired. The router is open-source and pinned in our repo — audit the logic yourself.
On Elite, every request is served from our EEA GPU and EU-resident model providers — your prompts never enter US discovery scope. Elite routes exclusively across EEA-GPU and EU-resident open-weight models. No US-hosted frontier models are provisioned or available on this tier. The default: nothing leaves the EU.
Refunds are paid in fiat (USD / EUR / SEK / NOK), not BTC. You receive the fiat value your crypto was worth on the day we recorded the inbound payment, minus per-second prorated usage. BTC price movement between purchase and refund is your exposure — we don't hedge FX. Refund window: cumulative usage < 3 hours across all orders ever (not calendar time). Fees shown in plain text before we send.
Pro+ accounts hold real spending power. A compromised account can drain credits faster than detection allows. FIDO2 (YubiKey, SoloKey, Apple/Google Passkeys) eliminates the phishing and credential-stuffing attack surface. Starter is FIDO2-optional; Pro / Elite require it at sign-in.
The YubiKey is FREE on Pro+ accounts. The cost is on us — not baked into per-token rates. Pick your delivery option at preorder:
Details in privacy §8a.
When public preorder volume crosses $3,500, we fund a second EU GPU node — expanding capacity and adding larger open-weight models to the Pro routing pool. Pro has always routed exclusively across our self-hosted + open-weight stack (Llama, DeepSeek, Mistral, Qwen, GLM); the threshold unlocks more GPU headroom. Progress is tracked on the homepage public counter.
Every fiat rail (Visa, Stripe, ACH, SEPA, SWIFT) puts a KYC-bearing intermediary between you and us. The no-KYC promise becomes structurally unenforceable the moment a single fiat payment clears our books. Crypto keeps that chain broken.
Longer take: see the Why we don't accept fiat callout above.
Preorder $20 → $26 in credits and you get a working key the moment the gateway opens (Mon 18 May 2026).
Pre-launch, the API base is https://api.llmdeal.me/v1 — run curl /v1/models
to see which models are live now. Model availability is public.