New lane: glm-5.3-1m on B200 — 114 tok/s, 1M context.TRY IT
Dedicated lanes · us-east · eu-central

Your model, in its own lane. Uncensored. Unlogged.

A dedicated serving lane per workload — profiled on your real requests, tuned to your latency target, retuned as traffic drifts. Open weights minus the refusals. Nothing stored. Billed per token.

Get an API key$curl api.runlane.us/v1
Throughput
412tok/s
TTFT p50
38ms
Refusal rate
0.4%
Retention
zero
NorthlightArbor HealthTesselQuillHexa LabsMeridianSableOrbit PayFathomKestrel
02 · Platformlanes · shadow · catalog · vault
Platform

Four things, done properly. Nothing you have to babysit.

Lanes
A serving stack per workload.

Batching, cache, quantization and decode strategy chosen for your request shape — and re-chosen when it drifts.

lane-01 · llama-3.3-70b · spec-decode412 tok/s
lane-02 · qwen3-235b · fp8288 tok/s
lane-03 · glm-5.3 · 1M-ctx114 tok/s
Shadow
Your traffic, measured before it’s served.

We mirror the first 10K requests and keep watching: latency percentiles, cache hit rate, output length — the inputs the tuner uses.

REQUESTS · 24H
87.3M
P50 / P99
38 / 840ms
PREFIX CACHE HIT
61%
RE-TUNES
3 this week
Catalog
Open weights as released. Minus the refusals.

Abliterated frontier models, no injected system prompt, no classifier in the output path. New weights land as lanes within days.

llama-3.3-70b-uncensoredopen lane →
qwen3-235b-uncensoredopen lane →
glm-5.3-1mopen lane →
gpt-oss-120b-uncensoredtuning · nov
Vault
Zero retention is the only mode.

Prompts live in GPU memory for one request. Dedicated lanes pin to a region and a single tenant. Audited quarterly.

retention0 bytes
logsmetadata only
regioneu-central-1 · pinned
tenancysingle
auditQ3 2026 · public ↗
03 · Catalog3 models · fp8 · 128K–256K ctx
Catalog

The weights, as released. Minus the refusals, plus a lane of their own.

Z.ai · GLM
● lane live
glm-5.3-1m
GLM-5.3 · 1M
The flagship lane. 753B MoE with a one-million-token window, served hot on B200.
Input
$0.75
per 1M tokens
Output
$1.50
per 1M tokens
Context
1M
tokens
Try glm-5.3-1m114 tok/s →
Alibaba · Qwen
● lane live
qwen3.8-27b
Qwen3.8 27B
The quick lane. Smallest, snappiest, cheapest token on the site.
Input
$0.10
per 1M tokens
Output
$0.20
per 1M tokens
Context
256K
tokens
Try qwen3.8-27b180 tok/s →
04 · Lanesrouter v2.4 · 5 active lanes
Lanes

A lane is a serving stack built around one workload.

Shared endpoints batch everyone the same way, so nobody gets the fast path. A lane is the opposite: batching, KV-cache, quantization and decode strategy chosen for the shape of your requests alone — and re-chosen when that shape changes.

01
Shadow
We mirror your first 10K requests and learn their shape: prompt length, output length, concurrency, cache hit rate.
02
Tune
Hundreds of serving configs — FP8/FP4, speculative decode, scheduler, kernels — are raced against your latency target on that traffic.
03
Hold
The winner goes live. When traffic drifts past a threshold the lane re-tunes itself — a changelog entry, never a page.
runlane router · live● 5 lanes
05 · Benchmarksus-east · 24h window · oct 2026
Benchmarks

Measured on a real traffic shape, not a leaderboard.

llama-3.3-70b on a shadowed production shape: 1K-in / 512-out, 64 concurrent streams, 24h from us-east. Methodology and raw traces are public — run them against your own lane.

Runlane412 tok/s
Provider A288 tok/s
Provider B241 tok/s
Provider C196 tok/s
TOKENS / S · HIGHER IS BETTER · OCT 2026
06 · Case study40M req / mo · p99 −52%
Case study · Arbor Health · 40M req / mo
p99 latency
840 ms
1.9 s · −56%
Cost / 1M tok
$0.61
$1.60 · −62%
Concurrency
3×
held, 0 re-configs
“We moved 40M requests a month onto one lane. Nobody on our side touched a config.”
Maya Okonkwo
CTO, Arbor Health
07 · Privacysoc 2 · hipaa · zero retention

Zero retention isn’t a tier. It’s the only mode we run.

Zero retention
Prompts live in GPU memory for one request.

Then they are gone. No request bodies in logs, no sampling “for quality”, no training set. Audited quarterly by a third party; the report is public.

Compliance
SOC 2 Type II. HIPAA BAA on request.

Dedicated lanes pin to one region and one tenant. eu-central lanes never leave Frankfurt — not for failover, not for support.

No refusals
0.4% refusal rate across the catalog.

Abliterated weights, no injected system prompt, no classifier in the output path. Your policy is the only policy on the lane.

08 · Integrationsopenai + anthropic compatible

Change one line. Keep every client, framework and agent you already run.

  • OpenAI SDKsdk
  • Anthropic SDKsdk
  • Vercel AI SDKsdk
  • Pythonsdk
  • TypeScriptsdk
  • cURLhttp
  • LangChainframework
  • LlamaIndexframework
  • Claude Codecli
  • Cursoride
  • n8nworkflows
  • OpenClawagent
// one line to switch
client = OpenAI(base_url="https://api.runlane.us/v1", api_key=RUNLANE_KEY)
09 · Pricingper token · no minimum
Pricing

Per token on shared lanes. Per hour on dedicated. Nothing else on the invoice.

INPUT TOKENS / MONTH120M
OUTPUT TOKENS / MONTH30M
MODEL
Estimated monthly
$135
The same tokens through a closed frontier API: $810. Difference: $675 a month.
Input120M × $0.75
Output30M × $1.50
Prompt caching−90% on hits
Commitmentnone
Dedicated lanes
single tenant · region pinned · hourly
B200 ×8180 GB HBM3e per GPU$2.90 /GPU-hrOpen lane →
H200 ×8141 GB HBM3e per GPU$2.10 /GPU-hrOpen lane →
H100 ×880 GB HBM3 per GPU$1.60 /GPU-hrOpen lane →
10 · Changeloglast deploy 2026-10-27 14:02 utc
Changelog

Lanes get faster while you sleep.

All releases →
2026-10-27llama-3.3-70b lanes re-tuned: 412 tok/sFused decode kernel on B200 lanes. +31% throughput at the same per-token price; 1,140 lanes re-tuned overnight.perf2026-10-19eu-central dedicated lanesSingle-tenant lanes pinned to Frankfurt. Requests, weights and logs-that-don’t-exist never leave the region.region2026-10-08Prefix caching on every lane90% off cached prefix tokens. On by default, nothing to change client-side.billing
11 · Startapi.runlane.us/v1

Open a lane.

One key, one base URL. The lane starts tuning itself from your first request.

Get an API keyTalk to an engineer