Skip to content

LLM INFERENCE · EU · FLAT RATE

The Meter is Dead. Long live the Token.

Open-weight models served in the EU region you choose, with pinned versions and zero prompt retention. Send as many tokens as your capacity carries; the invoice does not move.

Commons from €149/seat/month Demesne dedicated instances from €3,400/month

Video — coming soon

One Commons seat over twelve months: token volume varies by more than 20×, the invoice does not move.

Commons

From €149/seat/month

One seat per team member in a shared pool. Flat price per seat, unmetered tokens.

Demesne

From €3,400/month

Your own instance. Your own weights. Fixed monthly price, no one else on the hardware.

WHY KINGSLAYER

Predictable cost.

Per-token billing turns every product decision into a cost decision. One fixed monthly price keeps engineering and finance on the same plan.

Pinned versions.

Closed APIs can change model behavior under the same name. Kingslayer model versions are pinned until you choose to upgrade.

EU jurisdiction.

Data processed on US-controlled infrastructure is exposed to US legal process wherever the datacenter sits. Kingslayer processes exclusively in EU datacenters, under EU law.

PRICING

One price to rule them all.

Prices are public. The price on this page is the price on the order form, and it does not change during the contract term.

Commons

SEATS · 1 SEAT = 1 TEAM MEMBER

Volume pricing

1–7 seats €149 /seat
8–15 seats −10% €134 /seat
16–31 seats −15% €127 /seat
32+ seats −20% €119 /seat

Per seat

€134/seat/month

Monthly total

€1,072/month

Billed monthly or annually. The price is fixed for the term.

Get a quote

Every seat includes Unmetered tokens Guaranteed per-stream throughput ¹ Full model context Your pinned EU region Full model catalog

There is no free tier and no anonymous signup.

Demesne

D1

€3,400 /month

GPUs 1 × H100 80 GB
Fits ≤ 32B dense; small MoE
Throughput envelope Measured at onboarding ³
Tenancy Dedicated
Get a quote

D2

€5,900 /month

GPUs 2 × H100
Fits 70B-class dense
Throughput envelope Measured at onboarding ³
Tenancy Dedicated
Get a quote

D4

€10,900 /month

GPUs 4 × H100
Fits 70B long-context; mid-size MoE
Throughput envelope Measured at onboarding ³
Tenancy Dedicated
Get a quote

D8

On quote

GPUs 8 × H100
Fits Large MoE; frontier open weights
Throughput envelope Measured at onboarding ³
Tenancy Dedicated
Talk to a specialist

12-month minimum. Quarterly prepayment; annual prepayment −8%. One-time setup fee €1,500 covering onboarding load tests and integration. Resizes at quarter boundaries. Standby replica add-on available.

All prices in EUR, excluding VAT.

UNMETERED

Tokens are unmetered.

Send as many as your seat can carry. Your invoice is the same in your heaviest month and your quietest.

Every Commons seat and every Demesne instance includes every token its capacity can produce. There are no token tiers, no credits, no overage line, and no throttling to manage a bill. The usage graphs in your console are for your information. They are never a bill.

Token tiers
None.
Credits
None.
Overage billing
None, ever. Upgrades are more concurrency, more keys, more regions, better support.
Context window
The model's full context, on every plan.

INVOICE

Kingslayer · Bill to: your company

Period 2027-04 → 2027-06


Commons · 8 seats · 3 months

€3,216

Tokens generated in period · 471,220,904

not billed

Seats in use at peak · 8 of 8

included


Total due (excl. VAT)

€3,216

CONCURRENCY

We meter concurrency, not tokens.

A seat is a fixed number of concurrent requests with a guaranteed throughput per stream. Inside it, tokens are unmetered.

Concurrency is what makes the flat price possible: each seat comes with known parallel capacity, so your costs stay predictable while your traffic grows. Your console shows your headroom in real time, and when you need more, you simply move up a seat.

Within your cap
Guaranteed per-stream throughput: p50 ≥ 25 tokens/s on 70B-class models. ¹
Over your cap
Requests queue for up to 2 s, then receive HTTP 429 with a Retry-After header. Deterministic and documented.
Time to first token
p95 < 800 ms, in-region, within your cap. ¹
Need more
Add seats as your team grows; every seat carries the same guaranteed capacity. Volume pricing applies from 8 seats.

Video — coming soon

Concurrency in practice: eight streams at work, and exactly what happens over the cap.

DEDICATED · DEMESNE

Your own instance. No one else on the hardware.

Demesne is a dedicated inference instance: dedicated GPUs, dedicated hosts, a dedicated endpoint, run by Kingslayer as a managed service.

You control the three things a shared API never lets you control.

The model

Any compatible open-weight model: from the Kingslayer catalog, from the open ecosystem after license review, or your own fine-tunes and LoRA adapters.

The version

Pinned indefinitely. Upgrades happen when you schedule them, in your maintenance window, with a rollback path kept for 7 days.

The territory

Region selected at signature. The data plane never leaves it. Private connectivity available: IP allowlist, mTLS, WireGuard tunnel.

Tenancy
Dedicated GPUs and hosts. No shared inference process. No shared KV or prefix cache. No GPU slicing across customers, ever.
Control plane
Sees health, metrics, and configuration only. It never sees request content.
Capacity
A throughput envelope measured with your own traffic at onboarding, published to you as your numbers, and used as your SLO baseline.

DEMESNE · D2

DATA PLANE · SEALED

2 × H100 · dedicated GPUs and hosts

{name}.demesne.kingslayer.eu

model · pinned until you schedule an upgrade

region · fr-par-1 · the data plane never leaves it

prompts and completions stay inside

CONTROL PLANE

health

metrics

configuration

telemetry · outbound only

Only telemetry crosses the tenancy boundary. Request content never does.

FAIR USE

Fair by design.

Tokens are unmetered, and real production workloads never need a second thought. The policy comes down to common sense: your seats serve your own products, and the pool is shared fairly among everyone on it.

If the pool is under pressure, this is the order of operations. It is published, and it does not change without notice.

  1. 1 Burst capacity is provisioned.
  2. 2 Fair queuing tightens to guarantee every seat's floor.
  3. 3 The time-to-first-token target relaxes before the throughput target.
  4. 4 HTTP 503 with Retry-After, as a final state only.
A
8 / 8 at cap
B
2 / 2 at cap
C
3 / 8 headroom
D
1 / 8 headroom

A and B run at their full cap. C and D keep their headroom. No seat can slow down another — every seat keeps its guaranteed floor.

SOVEREIGN · OPEN

Your region. Your weights.

EU legal entity. EU datacenters. Open-weight models with published checksums. In writing, on every order form.

Sovereign

Region pinning
Each API key is pinned to a region at creation. Requests on that key are processed in that region only. There is no silent failover; cross-region failover is opt-in per key.
Zero retention
Prompts and completions are never written to disk. What we keep: timestamps, token counts, latency, model, and status, for operations and your dashboard.
Jurisdiction
Processing exclusively in EU datacenters, under EU law. GDPR DPA standard with every order form. Sub-processor list published and versioned.

Open

Open weights only
Every catalog entry publishes exact version, source, license, quantization, and the SHA-256 of its weights.
Immutable model IDs
An ID like qwen3-32b-fp8-2026a never changes what it points to. Aliases such as latest exist and are opt-in.
Deprecation
90 days written notice, a migration guide, and a side-by-side evaluation window.

API KEY

region fr-par-1
retention none
failover opt-in
dpa signed
sub-processors published

CATALOG ENTRY

model qwen3-32b-fp8-2026a
version pinned
license reviewed per model
quant fp8
sha256 published per entry

COMMONS OR DEMESNE

Which one do you need?

Commons

A seat in a shared pool.

Tenancy Shared pool, concurrency-capped seat
You choose Region, model from the catalog, number of seats
Best when Traffic is bursty, and region pinning, zero retention, and a DPA meet your compliance requirements
Price basis Per seat, per month
Live in Days

Demesne

Your own instance.

Tenancy Dedicated GPUs and hosts
You choose Region, any compatible open weights including your fine-tunes, hardware profile
Best when You need single-tenancy on paper, run custom weights, or sustain high load
Price basis Per instance, per month
Live in About two weeks, including your load test

Commons customers see in their console when a dedicated instance would serve them better, before any sales outreach.

INTEGRATION

Integration is a base URL change.

The API is OpenAI-compatible. Existing SDKs work unchanged.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.kingslayer.eu/v1",
    api_key=os.environ["KINGSLAYER_API_KEY"],
)

stream = client.chat.completions.create(
    model="qwen3-32b-fp8-2026a",
    messages=[{"role": "user", "content": "Summarize this clause."}],
    stream=True,
)
Endpoints
POST /v1/chat/completions, streaming and non-streaming. GET /v1/models.
Structured output
JSON-schema-guided decoding. Tool calling where the model supports it.
Response headers
Every response carries its region and its model pin, plus your current concurrency headroom.
Errors
OpenAI-shaped error envelopes. Over-cap requests return 429 with Retry-After.

FAQ

Frequently asked questions.

What does "unmetered" mean?

Inside your seat or instance, tokens are not counted toward your bill. There are no token tiers, credits, or overage charges. The limit is concurrency: how many requests run at once.

What happens when I exceed my concurrency cap?

Requests above your cap queue for up to 2 seconds. If a stream does not free up, the request receives HTTP 429 with a Retry-After header. Nothing is billed, nothing is dropped silently, and your console shows your headroom in real time.

Where is my data processed, and what do you keep?

In the EU region pinned to your API key, and nowhere else unless you opt in to cross-region failover. Prompts and completions are never written to disk. We keep request metadata: timestamps, token counts, latency, model, and status.

Which models can I use?

Commons serves an open-weight catalog; every entry publishes its version, source, license, quantization, and SHA-256. Demesne runs any compatible open-weight model, including your own fine-tunes and LoRA adapters, after license review.

Do model versions change without notice?

No. Model IDs are immutable. You upgrade when you decide. Deprecations come with 90 days written notice, a migration guide, and a side-by-side evaluation window.

How do I migrate from another OpenAI-compatible API?

Change the base URL and the model ID. Chat completions, streaming, structured output, and tool calling use the same request and response shapes.

Is there a free tier or self-serve signup?

No. Kingslayer sells to businesses under a signed order form.

How are contracts and payments structured?

Commons is billed per seat, monthly or annually, with volume pricing from 8 seats. Demesne profiles are annual contracts with quarterly prepayment and offer −8% for annual prepayment. Prices are in EUR, excluding VAT, and do not change during the term.

Who is behind Kingslayer?

An EU-based company. Processing is performed exclusively in EU datacenters; the GPU providers are listed as sub-processors in the DPA.

Unmetered. Sovereign. Open.

Tell us your traffic shape and your compliance requirements. You receive a quote and a draft order form, with every contract document available before signature.

  1. Throughput and time-to-first-token targets are measured at the gateway, in-region, within the seat’s concurrency cap, on 70B-class models served at FP8. Smaller models are faster; the targets are floors, not typical values.
  2. Prices in EUR, excluding VAT. Commons is billed per seat, monthly or annually, with volume pricing from 8 seats. Demesne profiles are annual contracts with quarterly prepayment. Prices are fixed for the contract term and are identical here and on the order form.
  3. Demesne throughput envelopes depend on model, context length, and traffic shape. They are measured with the customer’s own prompts during onboarding, published to the customer, and used as the SLO baseline.
  4. Standby replica: a second instance in the same region for faster failover, priced at 60% of the profile price. The setup fee covers onboarding load tests and integration support.