SABLE
DocsOpen Console
AI Inference Cloud · v3

Run frontier models at the speed of one API call.

Sable serves open and custom LLMs on a single OpenAI-compatible endpoint: sub-250ms first token, autoscaling to zero, and per-token billing. No cluster to babysit, no cold-start tax.

Start building, free $25 credit Browse models
$ pip install sable · deploy in under 60 seconds · no card required
GPU servers in a data center rack
POST /v1/chat/completionsllama-3.1-70b
# stream a 70B model, OpenAI-compatible
from sable import Sable
client = Sable(api_key="sk-live-…")

stream = client.chat.completions.create(
  model="llama-3.1-70b",
  messages=[{"role":"user",
    "content":"Explain KV-cache reuse."}],
  stream=True,
)
for chunk in stream:
  print(chunk.delta, end="")
First token184ms p50
Throughput1,850 tok/s
Uptime99.99%

Powering inference for

42Btokens / day
68served models
6global regions
<250ms p50 TTFT
0day data retention
01 · Model catalog

One endpoint. Every open model.

Serverless and warm. Swap models by changing a string; pay only for the tokens you generate. Prices below are per 1M output tokens.

Llama 3.1 70B

meta · general reasoning
$0.60$ / 1M out
chat
Context 128K · TTFT 184ms · Throughput 1,850 t/s

Llama 3.1 8B

meta · low-latency workhorse
$0.09$ / 1M out
fast
Context 128K · TTFT 86ms · Throughput 4,200 t/s

Mixtral 8×22B

mistral · mixture-of-experts
$0.90$ / 1M out
moe
Context 64K · TTFT 232ms · Throughput 1,120 t/s

Qwen2.5 72B

alibaba · code + reasoning
$0.70$ / 1M out
multilingual
Context 128K · TTFT 221ms · Throughput 1,610 t/s

DeepSeek-V3

deepseek · frontier open weight
$1.10$ / 1M out
671B moe
Context 128K · TTFT 318ms · Throughput 940 t/s

Sable-Embed v2

sable · 1024-dim retrieval
$0.02$ / 1M out
embeddings
Context 8K · TTFT 38ms · Throughput 12,400 t/s

Bring your own weights: upload a fine-tune or LoRA adapter and Sable serves it on the same API within 90 seconds. Benchmarks measured at 512-token prompts, us-east-1, batch 1.

02 · Architecture

From request to token, engineered.

A speculative-decoding runtime, continuous batching, and paged KV-cache keep GPUs saturated so your latency stays flat under load.

/01

Point your SDK at one URL

Drop-in OpenAI compatibility. Change the base URL and key; your existing client libraries, tools, and agents work unchanged.

/02

We route to the warmest GPU

Requests land on the nearest region with capacity. Speculative decoding drafts ahead while continuous batching packs the accelerator.

/03

Tokens stream back in milliseconds

Paged KV-cache reuses shared prefixes across requests, so repeated system prompts and RAG contexts cost you almost nothing.

/04

Scale to zero, or pin dedicated

Serverless idles down between bursts. Need guaranteed headroom? Reserve dedicated H100s with per-second billing and no cold starts.

3.4×

throughput vs. vanilla vLLM

  • Speculative decoding + Medusa heads
  • Continuous batching, no head-of-line stalls
  • Paged & prefix-shared KV-cache
  • FP8 + AWQ quantized serving paths
  • Streaming, JSON-mode & tool-calling native
03 · Pricing

Pay for tokens, not idle GPUs.

Start serverless and per-token. Graduate to reserved GPUs when your traffic is steady. No minimums, no egress fees within-region.

Serverless

$0.09/ 1M tokens

Per-token billing on shared capacity. Scales to zero automatically between requests.

  • 68 open models, one endpoint
  • 600 requests / min default limit
  • Community Slack support
Start free

Dedicated

most popular
$2.90/ H100·hr

Reserved accelerators, per-second billing, zero cold starts. A100 from $1.40/hr.

  • Guaranteed throughput & latency SLA
  • Custom fine-tunes & LoRA hosting
  • Up to 40% off on 1-yr reserve
Reserve capacity

Enterprise

Custom

Private VPC deployments, dedicated regions, and a 99.99% contractual uptime SLA.

  • Single-tenant & on-prem options
  • SOC 2 report & DPA / BAA
  • Named solutions engineer
Talk to sales
04 · Global footprint

Inference close to your users.

Six regions on NVMe-backed GPU fleets. Round-trip figures are intra-region p50, measured from a co-located client.

Network cabling in a GPU fleet
us-east-1N. VirginiaLive11ms
us-west-2OregonLive13ms
eu-central-1FrankfurtLive12ms
eu-west-1DublinLive14ms
ap-northeast-1TokyoLive15ms
ap-south-1MumbaiLive16ms
05 · Trust & compliance

Your prompts are yours. We keep nothing.

Zero data retention on inference by default: requests and completions are never logged or used for training. Encryption in transit and at rest, scoped API keys, and audited access on every path.

Attestation

SOC 2 Type II

Audited annually · report on request

Standard

ISO 27001

Information security certified

Healthcare

HIPAA-ready

BAA available on Enterprise

Data

Zero retention

No training on your traffic, ever

Ship your first inference before your coffee cools.

Create a key, paste it into any OpenAI client, and stream tokens in the next five minutes. $25 in credit, no card, no sales call.

Get your API keyRead the docs