Every model discloses its quantisation

The inference provider that tells you what you’re actually running

OpenAI-compatible endpoints for open-weight models. We publish the exact quantisation served, the hardware behind it, live latency and 30-day uptime — for every model, on a page anyone can read.

Your first top-up gets 10% extra credit.

quickstart.pypython
from openai import OpenAI

client = OpenAI(
    base_url="https://api.infersia.com/v1",
    api_key="isk-v1-...",
)

stream = client.chat.completions.create(
    model="qwen/qwen3-8b",
    messages=[{"role": "user", "content": "Explain KV caching."}],
    stream=True,
)

for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

Change two lines. The official openai SDKs work unmodified — streaming, tools, and all.

Serving right now

Latency and uptime measured from real traffic, not a benchmark.

Full catalogue →
S
Step 3.7 Flash

stepfun-ai/step-3.7-flash

Live
Quant
NVFP4
Context
256K
p50 TTFT
102ms
Uptime 30d
100.0%
B200
$0.18 in/$1.05 out
Qwen logo
Qwen3.6 35B A3B

qwen/qwen3.6-35b-a3b

Live
Quant
AWQ int4
Context
256K
p50 TTFT
231ms
Uptime 30d
99.8%
RTX PRO 6000 WS
$0.10 in/$0.90 out
Qwen logo
Qwen3 14B

qwen/qwen3-14b

Live
Quant
AWQ int4
Context
128K
p50 TTFT
54ms
Uptime 30d
99.9%
RTX 5090
$0.09 in/$0.22 out
Qwen logo
Qwen3 8B

qwen/qwen3-8b

Live
Quant
AWQ int4
Context
32K
p50 TTFT
195ms
Uptime 30d
94.9%
RTX 3090
$0.05 in/$0.15 out

Most providers won’t tell you the quantisation

It’s the number that determines whether the model you’re paying for behaves like the model you benchmarked. We treat hiding it as a bug.

Quantisation, always disclosed

Every endpoint states its precision as a first-class field — fp8, awq-int4, mxfp4 — alongside the engine and version. It’s in the catalogue, in /v1/models, and in a response header on every request.

Real numbers, not marketing ones

Latency and throughput come from your traffic and ours, measured continuously and published per model. Uptime is computed from health probes we keep on record — including the days it wasn’t 100%.

Zero prompt retention, structurally

There is no column in our database a prompt could go in, and the logger drops any field not on an allowlist. It isn’t a policy we follow — it’s a shape the system has.

Works with what you already use

We implement the OpenAI API surface exactly — request shapes, response shapes, streaming, error types and rate-limit headers. If a tool supports a custom OpenAI base URL, it supports us.

  • openai-python & openai-nodeOfficial SDKs, unmodified
  • Cline, Continue, AiderPoint at our base URL
  • LangChain, LlamaIndex, CrewAIChatOpenAI-compatible
  • LiteLLM, OpenWebUIStandard provider config
curlbash
curl https://api.infersia.com/v1/chat/completions \
  -H "Authorization: Bearer $INFERSIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3-8b",
    "messages": [{"role": "user", "content": "Hello"}],
    "stream": true
  }'
response headershttp
x-request-id:              isr_01KYGHW7S1GTR2AVE40C
x-infersia-provider:      infersia
x-infersia-quantization:  awq-int4
x-infersia-cost-usd:      0.000001950

Every response tells you what served it and what it cost — to the nano-dollar.

Get 10% on your first top-up

Sign up, get a key, and point your existing code at a real endpoint. Your first top-up lands with 10% extra credit — $10 minimum, no subscription, no expiry.