v2 · streaming inference, live

Inference at the edge
of latency.

Helix is the inference API for agents that can't wait. Sub-40 ms first token, streamed from 312 edge nodes — one endpoint, every frontier model.

Start building See the benchmarks $5 free credits · no card
stream.ts ● ttft 38ms
// one endpoint, every model
const res = await helix.stream({
  model: "helix/auto",
  route: "nearest-edge",
  input: agent.turn,
  budget_ms: 40
})

for await (const tok of res) emit(tok)
40ms ceiling
Shipping in production at
Northwind Cael AI Drift&Co Loom.dev Polaris Brightloop
Latency is the product

Other APIs make you wait. Helix answers before they finish thinking.

We routed a million agent turns through Helix and the median provider. Same models, same prompts. The gap is the part your users feel.

38ms
p99 time-to-first-token, measured at the edge
312
edge nodes across 41 metros, auto-routed
6.4×
faster cold-start vs. the median hosted API
99.99%
uptime, single SLA across every model
The platform

One key. Every model. Zero rewrites.

Swap a base URL and you're done. Helix speaks the chat-completions dialect you already wrote against — then adds the parts the spec forgot.

Smart routing, on by the token

`helix/auto` picks the cheapest model that clears your quality bar, then drops to a faster one mid-stream if a turn is simple. You set a latency budget; we hold the line.

route: nearest-edge · budget_ms: 40 01

Streaming that never stalls

True token-level SSE with mid-flight cancel and resume. Kill a runaway generation without burning the rest of your budget.

SSE · cancel() · resume() 02

Zero-retention by default

Prompts and completions are dropped the instant a request closes. SOC 2 Type II, no training on your data — ever, on every tier.

soc2 · zdr · region-pinned 03

Spend you can actually read

Per-agent, per-route cost telemetry streamed to your dashboard in real time — not a CSV that lands three days into the next billing cycle. Set hard caps; we stop the bleed before it's an incident.

live cost stream · per-agent caps · alerts 04
The network

Compute that lives where your users do.

Helix runs inference on 312 nodes pressed up against the backbone. Requests never hairpin to a single region — they resolve at the metro nearest the caller.

SFO · 9ms FRA · 11ms SIN SYD
  • 01
    Anycast on the request, not the region
    DNS resolves every call to the nearest healthy node. No config, no region picker, no 2 a.m. failover page.
  • 02
    KV-warmed model weights
    Hot models stay resident at the edge, so first-token latency doesn't include a cold container spin-up.
  • 03
    Region-pinned when you need it
    Compliance turn the dial: pin a workload to EU-only or US-only nodes and Helix routes inside the fence.

Give your agents their reflexes back.

Spin up a key, point your base URL at Helix, and watch your first-token times collapse. $5 in credits on signup — enough to run the benchmark yourself.

Published with shiply →