Skip to content
Concept demo

A fictional brand, designed and built by Get Ranked On Top. Not a real company.

Skip to content

All systems nominal · 14 of 14 regions serving

Inference at the speed of thought.

112 ms median time-to-first-token. 1,184 tokens per second on a 70B model. 60+ open-weight models behind one API, on GPUs we schedule ourselves across 14 regions — serverless, dedicated or inside your VPC.

41.7B
tokens served / day
112 ms
p50 time to first token
99.993%
API uptime, trailing 90 d
14
regions, 3 continents of GPUs

In production at

HALDEN/LABStidewaterParallax▲LegalMOSAIC HEALTHquorum.aiLattice & LoomNORTHFORMbrevet

01Playground

Watch the first token land.

Pick a prompt. The readouts show what our edge measures on every request: time to first token, then sustained throughput.

  • Speculative decodingon by default
  • Continuous batching256 seq / GPU
  • KV-cache reuse75% off cached input
  • Cold start, serverless0 ms · always warm
playground · api.halvorn.example/v1ready
Choose a preset prompt to run it on a live-shaped endpoint.
// output streams here
// readouts are measured from request start
model
—
time to first token
0 ms
throughput
0 tok/s
tokens
0

02Model catalogue

62 models. One endpoint.

Every model ships with tool calling, JSON mode and streaming. New open-weight releases go live within 48 hours, benchmarked before they are listed.

Halvorn model catalogue with context length, price per million tokens and latency
ModelTypeContextInput $/MOutput $/Mp50 TTFTThroughputStatus
vela-3-405bchat128k$3.20$3.20214 ms412 tok/sserving
vela-3-70bchat128k$0.54$0.72112 ms1,184 tok/sserving
vela-3-8bchat128k$0.06$0.0871 ms2,940 tok/sserving
orrin-coder-34bcode64k$0.38$0.48131 ms1,026 tok/sserving
lumen-vl-90bvision32k$0.90$1.10188 ms604 tok/sserving
tern-reason-32breasoning128k$0.42$1.68146 ms880 tok/sserving
sable-embed-2embed8k$0.04—18 ms—serving
sable-rerank-2rerank8k$0.05—24 ms—serving

Showing 8 of 62. Latency: client-measured, same-metro, 512-token prompt, 30-day p50.

03Benchmarks

2.4× the throughput. 61% less wait.

vela-3-70b, identical weights and sampling settings on every provider. We publish the harness; run it yourself before you believe us.

Output throughput · vela-3-70bhigher is better
  • Halvorn1,184 tok/s
  • Provider A486 tok/s
  • Provider B352 tok/s
  • Provider C301 tok/s
  • Self-hosted, 8-GPU node238 tok/s
Time to first token · p50lower is better
  • Halvorn112 ms
  • Provider A284 ms
  • Provider B391 ms
  • Provider C446 ms
  • Self-hosted, 8-GPU node612 ms
Workload512 in / 1,024 out tokens, 32 concurrent streams
Window72 h, 1.2M requests, Sep 2026
Clientsame metro as endpoint, HTTP/2 keep-alive
Providersanonymised, public list pricing tier

04Quickstart

Change one URL. Keep your code.

The SDKs follow the chat completions schema you already ship. Most migrations are a base URL and a model string.

  1. Create a keyScoped per project, with a hard monthly spend cap.
  2. Point your clientapi.halvorn.example/v1, or install the SDK.
  3. StreamServer-sent events, tool calls and JSON mode out of the box.
stream.py
from halvorn import Halvorn

client = Halvorn()  # reads HALVORN_API_KEY

stream = client.chat.completions.create(
    model="vela-3-70b",
    messages=[{"role": "user", "content": "Ship it."}],
    max_tokens=512,
    stream=True,
)

for chunk in stream:
    print(chunk.delta, end="", flush=True)

05Regions

14 regions. Under 20 ms from 91% of the internet.

Anycast routing sends every request to the nearest healthy GPU pool. Pin a key to a region, or to EU-only, for data residency.

Map of Halvorn's 14 GPU regions across North America, South America, Europe, Asia and Oceaniasfo1pdx1ord1yyz1iad1gru1lhr1fra1arn1bom1sin1icn1hnd1syd1
Regions with median round-trip time from the local metro
RegionMetroRTT
sfo1San Francisco4 ms
pdx1Portland11 ms
ord1Chicago9 ms
yyz1Toronto12 ms
iad1Ashburn7 ms
gru1São Paulo18 ms
lhr1London6 ms
fra1Frankfurt5 ms
arn1Stockholm9 ms
bom1Mumbai14 ms
sin1Singapore8 ms
icn1Seoul10 ms
hnd1Tokyo6 ms
syd1Sydney13 ms

06Pricing

Per million tokens. Nothing else.

No seat fees, no minimums on serverless, no egress charges. Cached input is 75% off; the batch API is 50% off with a 24-hour window.

Serverless price per million tokens in US dollars
ModelInputCached inputOutputBatch input
vela-3-8b$0.06$0.015$0.08$0.03
vela-3-70b$0.54$0.135$0.72$0.27
vela-3-405b$3.20$0.80$3.20$1.60
orrin-coder-34b$0.38$0.095$0.48$0.19
tern-reason-32b$0.42$0.105$1.68$0.21
lumen-vl-90b$0.90$0.225$1.10$0.45
sable-embed-2$0.04——$0.02

Serverless

$0

to start · pay per token

Shared, autoscaled endpoints for every catalogue model. Rate limits rise automatically with spend.

  • $5 free credit, no card
  • 600 RPM → 60,000 RPM
  • Batch API at 50% off
  • Community support
Get API key

Dedicated

most chosen

$2.85

per GPU-hour · billed per second

Reserved capacity with your own weights, LoRA adapters and a private endpoint in any region.

  • Bring your own fine-tune
  • Autoscale 1 → 512 GPUs
  • 99.95% uptime SLA
  • Shared Slack channel
Reserve capacity

Enterprise

Custom

annual commit

Halvorn's runtime in your VPC or data centre, with committed-use pricing and a named engineer.

  • VPC and on-prem deploys
  • Zero retention, contractually
  • 99.99% uptime SLA
  • 15-min P1 response
Talk to an engineer

07Enterprise

Your VPC. Your keys. Zero retention.

The same runtime that serves 41.7B tokens a day, deployed where your data already lives. Control plane with us, data plane with you.

VPC

Runs in your cloud account

Our scheduler and kernels deploy into your VPC via a single Terraform module. Weights, prompts and logs never cross your boundary.

0 s

Zero data retention

Prompts and completions are held in memory for the life of the request, then dropped. No training on your traffic, ever — it is in the contract.

BYOK

Your keys, your audit trail

Customer-managed encryption keys, SSO with SCIM, per-key spend caps and an immutable audit log exported to your SIEM.

1:1

A named inference engineer

Capacity planning, kernel tuning for your sequence lengths, and a 15-minute P1 response window around the clock.

08Customers

Measured in production.

From 3-person startups to teams pushing 2B+ tokens a day. Numbers are theirs.

“We moved 2.3 billion tokens a day across in a weekend. p95 latency on our support agent fell from 2.1 s to 640 ms, and the invoice fell by 41%.”
Mira CastellanosHead of Platform, Tidewater
“Zero retention was the gate for our legal team. Halvorn put it in the MSA on the first call instead of the fourth.”
Jonah LindqvistCTO, Parallax Legal
“Speculative decoding on our own fine-tune, in our VPC, without us writing a CUDA kernel. That was the whole pitch, and it held.”
Priya RamanStaff ML Engineer, Halden Labs

09Talk to an engineer

Bring a workload. Leave with a number.

Tell us your model, peak RPM and sequence lengths. We reply with a latency and cost estimate from a real benchmark run — usually inside a day.

  • Salessales@halvorn.example
  • Supporthelp@halvorn.example
  • Phone+00 415 010 2266
  • HQ480 Harrow St, Floor 6, San Francisco
Reply in 1 business day. No sales sequence.

10FAQ

Questions engineers ask first.

Is the API a drop-in replacement for what we use today?

Halvorn implements the de-facto chat completions, embeddings and batch schemas. Most teams change a base URL and a model string; streaming, tool calls and JSON mode behave identically.

How are latency numbers measured?

p50 time-to-first-token is measured client-side from a request leaving a VM in the same metro to the first streamed byte, 512-token prompts, 30-day rolling window. We publish the raw percentiles daily at status.halvorn.example.

Do you train on our prompts or outputs?

No. Serverless traffic is retained for 0 seconds by default; opt-in 30-day logging exists only for your own debugging and is encrypted with your key.

Can we run our own fine-tuned weights?

Yes. Upload safetensors or point us at a bucket. Dedicated endpoints serve full fine-tunes; serverless endpoints hot-swap up to 1,000 LoRA adapters per base model at no extra cost.

What happens if a region degrades?

Requests fail over to the nearest healthy region in under 3 seconds unless you pin residency. EU-pinned keys only ever fail over between lhr1, fra1 and arn1.

How does billing work?

Per million tokens on serverless, per GPU-second on dedicated, both invoiced monthly. Cached input tokens are billed at 25% of the input rate; the batch API is 50% off with a 24-hour window.