Inference that scales to zero
Deploy vLLM or any container as an autoscaling endpoint. Sub-300 ms p95 under load, zero replicas when traffic sleeps, and billing that stops the millisecond the response does.
Pay per request-second · OpenAI-compatible API · Scale 0 → 64 · SOC 2 Type II
Inference playground
contract-summarizer.reviosa.app/v1Summarize the change-of-control language in this MSA and flag anything non-standard.
Section 11.2 is a standard mutual change-of-control trigger — either party may terminate on 30 days’ notice following an acquisition.
Flag: the assignment carve-out in §11.4 survives termination, which is unusual for a vendor-side MSA. Suggested redline below…
TTFT 178 ms · 412 tok/s · billed 0.62 request-seconds
Your model, a URL, and nothing to babysit
Any open-weight model or container becomes an autoscaling HTTPS endpoint — tuned vLLM underneath, observability on top, zero idle cost in between.
New endpoint
vLLMHugging Face model
meta-llama/Llama-3.1-8B-Instruct
Custom container
ghcr.io/quillquery/reranker:v4
GPU class
H100 SXM · 80 GB
Deploy any model
Point at a Hugging Face repo and we serve it on tuned vLLM — paged attention, FP8, continuous batching, already dialed in. Or bring any OCI container that answers HTTP; if it has a port, it can be an endpoint.
Replicas · last 24 h
now: 0 · $0.00/hrAutoscale to zero
Replicas follow queue depth, not a schedule. Spikes scale 0 → 64 in seconds; when requests stop, so does the meter. Idle weekends bill exactly $0.00.
contract-summarizer
Last hourp95
287 ms
Decode
412 tok/s
Errors
0.02%
200 POST /v1/chat/completions · 178 ms ttft · 512 tok
200 POST /v1/embeddings · 24 ms · 96 tok
trace 9f2c81 exported · otlp · us-east
Observability built in
Every request is traced: time-to-first-token, tokens per second, GPU utilization, cost. Stream logs from the console or pipe metrics to your own stack — no sidecar to install.
Zero to endpoint in one curl
The whole lifecycle is an API — create, deploy, scale, observe. And the URL you get back speaks /v1/chat/completions, so your existing OpenAI SDK client works by changing one line.
Get your API key-
1
Declare the endpoint.
Model or container image, GPU class, scale bounds. One POST — no YAML, no Helm charts, no operators to install.
-
2
We roll it out.
A tuned vLLM image lands on the fleet and health checks gate the cutover — about 40 seconds for an 8B on H100 SXM. Rollouts are atomic; rollbacks are one call.
-
3
Point your SDK at it.
Swap the base URL and keep every other line of client code. Streaming, tool calls, and embeddings all behave the way your libraries expect.
$ curl -s https://api.reviosa.com/v1/endpoints \
-H "Authorization: Bearer $REVIOSA_API_KEY" \
-d '{
"name": "contract-summarizer",
"engine": "vllm",
"model": "meta-llama/Llama-3.1-8B-Instruct",
"gpu_class": "h100-sxm",
"scale": { "min": 0, "max": 16 }
}'
{
"id": "ep_7f3d92",
"status": "deploying",
"url": "https://contract-summarizer.reviosa.app/v1",
"gpu_class": "h100-sxm",
"billing": "per-request-second"
}
$ reviosa endpoints watch ep_7f3d92
✓ image pulled · vllm-0.9-fp8 .......... 18.2 s
✓ health check passed · us-east ....... 21.4 s
live — https://contract-summarizer.reviosa.app/v1POST to live URL
~40 s
8B model, H100 SXM class
Client code changed
1 line
the base URL — OpenAI-compatible
Fast enough for the request path
Serverless usually means a slow first hit. We keep warm pools and pinned weight caches in every region, so scale-from-zero feels like scale-from-one.
287 ms
p95, end to end
512-token completions, Llama-3.1-8B, us-east
1.9 s
cold start
zero to first token on a scale-from-zero request
412 tok/s
per-replica decode
H100 SXM, FP8, continuous batching
99.95%
endpoint availability
trailing 90 days, all four regions
Measured nightly from client VMs in-region against production endpoints — medians of 1,000 runs, never synthetic benchmarks.
GPU-backed
Serverless in front, H100s underneath
Endpoints run on the same Reviosa-owned fleet our customers rent by the second — not a mystery pool. You choose the silicon per endpoint; we keep it warm across four regions.
- Pick the class per endpoint — 80 GB H100 SXM by default; step up to H200’s 141 GB for 70B-class models, or down to L40S and RTX 4090 for embeddings and rerankers.
- Whole GPUs, single tenant — your replicas never share silicon with another customer, and VRAM is wiped between scale events.
- A path out of serverless — the same container runs unchanged on rented GPU instances when you want the whole card, the whole node, or the whole rack.
contract-summarizer · Hardware
us-eastH100 SXM
Default80 GB HBM3 · FP8 transformer engine
H200 SXM
70B+141 GB HBM3e · long-context serving
L40S
48 GB GDDR6 · small models, vision
RTX 4090
Embeddings24 GB GDDR6X · lowest cost per request
Metered per request-second against the class you choose.
Compare GPUsMargaret Liu
VP of Engineering @ Quill & Query
“Contract review traffic lives between 9 and 6 — our old cluster idled sixteen hours a day. On Reviosa the endpoints scale to zero at night and we simply stop paying. And with p95 under 300 ms, clause summaries render inline in the review screen instead of in an overnight batch.”
61%
lower monthly inference spend after replacing always-on GPU servers with scale-to-zero endpoints.
Read customer storiesStart training in minutes
Create an account, add a card, and launch your first GPU instance. Per-second billing means you only pay for what you use.
No minimum commitment · Cancel anytime · $10 free credit for new accounts