Skip to main content

Inference.net Provider Guide

Inference.net is a Tier-2 catalog provider: OpenAI-wire-compatible with no behavioural quirks, so its entire integration is one JSON file (src/lib/providers/catalog/inference-net.json) rather than hand-written code. That file is the source of truth for everything on this page.

Verification status: this entry is docs- and roster-verified, not yet live-verified. It was onboarded credential-free — built entirely from the vendor's public documentation and an unauthenticated GET /v1/models call, with no API key ever created or used. No chat/stream/tool/structured-output request has been made against the live API yet. evidence.liveMatrix is null until someone runs the live capability sweep in Live verification with a real key.

Key Facts​

  • Provider id: inference-net
  • Protocol: OpenAI-compatible (/chat/completions), Authorization: Bearer <key>
  • Base URL: https://api.inference.net/v1
  • Default model: glm-5.2 — the exact model used in the vendor's own API quickstart code sample
  • Models in catalog: 8 (curated slice of a live 63-model roster spanning Claude, GPT, Gemini, DeepSeek, GLM, Kimi, Llama, Qwen, Gemma, Nemotron and Inference.net's own Schematron models)
  • Streaming: documented (stream: true in the quickstart example)
  • Tool calling: documented (tools request parameter, "function definitions for function calling") — not live-probed
  • Structured output: documented (response_format accepts json_object or a JSON schema) — not live-probed
  • Embeddings: not documented
  • Billing: free-tier — pay-as-you-go, $0-usage entry tier (1M gateway requests/month, 30 req/min)
  • Key format: none declared
  • Env var: INFERENCE_API_KEY (the vendor's own docs use this exact name in their quickstart code, which is why the catalog sets wire.envOverrides instead of the INFERENCE_NET_API_KEY name NeuroLink's naming convention would otherwise derive from the inference-net provider id)

Quick Start​

1. Get an API key​

  1. Visit: https://inference.net/register/ and sign up
  2. Create an Inference.net project API key from the dashboard (this is the gateway key — do not confuse it with a third-party provider key used only for the separate "proxy mode")
  3. New accounts start on a pay-as-you-go, $0-usage tier; confirm current terms at https://inference.net/pricing before heavy use
  4. Set INFERENCE_API_KEY in your .env file

2. Configure​

export INFERENCE_API_KEY=your-api-key
export INFERENCE_NET_MODEL=glm-5.2                        # optional — overrides the default model
export INFERENCE_NET_BASE_URL=https://api.inference.net/v1 # optional — self-hosted or proxy

3. Use it​

import { NeuroLink } from "@juspay/neurolink";

const neurolink = new NeuroLink();

const result = await neurolink.generate({
  input: { text: "Explain context windows in one paragraph." },
  provider: "inference-net",
  model: "glm-5.2",
});

console.log(result.content);
# CLI
npx @juspay/neurolink generate "Hello" --provider inference-net

Per-request credentials work as they do for every provider:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "inference-net",
  credentials: { "inference-net": { apiKey: process.env.INFERENCE_API_KEY } },
});

Models​

ModelContextVision$/M in · out · cacheNotes
glm-5.2 ⭐1Mno$1.30 / $4.40 (cached $0.20)Recommended — used in the vendor's own quickstart request example; native Zhipu GLM-5.2, hosted directly by Inference.net.
glm-5.31Mno$0.90 / $3.00 (cached $0.15)Newer GLM generation; one of the few roster models whose live entry lists chat + responses + messages endpoint support.
glm-5.3-flash1Myes$0.09 / $0.28 (cached $0.02)Cheapest model in this entry; the only one of these 8 whose live roster entry documents image + video input.
llama-3.3-70b-instruct131Kno$0.13 / $0.40Meta Llama 3.3 70B Instruct, open-weight; widely-recognized baseline fallback.
deepseek-v3.2164Kno$0.70 / $2.00 (cached $0.35)DeepSeek V3.2, open-weight fallback.
gpt-4.1-mini1Myes$0.40 / $1.60 (cached $0.10)OpenAI GPT-4.1 mini, reachable through Inference.net's gateway.
gemini-2.5-flash1Myes$0.30 / $2.50 (cached $0.03)Google Gemini 2.5 Flash, reachable through the gateway.
claude-haiku-4-5200Kyes$1.00 / $5.00 (cached $0.10)Anthropic Claude Haiku 4.5, reachable through the gateway; Claude-family entries list only the "messages" endpoint.

Context, pricing and vision (input_modalities) for every row above were read directly from the live, unauthenticated GET https://api.inference.net/v1/models response on 2026-09-28 (HTTP 200, 63 models total on the roster). This catalog entry lists a curated 8-model slice of that roster, not all 63 — the full list spans Claude, GPT, Gemini, DeepSeek, GLM, Kimi, Llama, Qwen, Gemma, Nemotron and Inference.net's own Schematron structured-extraction models.

Fallback order when the default is unavailable: glm-5.3 → glm-5.3-flash → llama-3.3-70b-instruct → deepseek-v3.2 → gpt-4.1-mini → gemini-2.5-flash → claude-haiku-4-5.

Verification status​

Tier-2 onboarding requires evidence before a provider is accepted, and pnpm run verify:provider-onboarding gates it in CI. This is what the catalog records for Inference.net today:

ProbeResult
Rosterunauthenticated GET /v1/models, HTTP 200, 63 models, 2026-09-28. No API key was created or used.
Auth rejectionNot probed — a real or invalid-key request would be an authenticated/POST call, which credential-free onboarding does not make. errorRules declares a conventional HTTP 401 → authentication mapping based on the vendor's "OpenAI-compatible" framing, not a live 401.
Capability probesNot run. tools and response_format are asserted only because the vendor's own API reference documents them as request parameters (https://docs.inference.net/api/api-quickstart, retrieved 2026-09-28); no combined tools+schema request was ever sent, so structuredOutputWithTools and toolsWithStreaming are both false.
Live capability sweepNot run. evidence.liveMatrix is null. A future pass with a real key must run the live matrix per Live verification before this line is filled in.

Because no live request has ever been sent, every capability flag above should be read as "documented by the vendor," not "proven to work." NeuroLink's generic conflict-retry and JSON-repair layers are the intended safety net if a documented capability turns out to behave differently once probed live.

Troubleshooting​

SymptomCauseFix
Invalid Inference.net API keyINFERENCE_API_KEY unset or wrongCheck the key in the Inference.net dashboard: https://inference.net/register/
Model not foundThe roster changed since 2026-09-28Pick a current id from the live GET /v1/models roster
Key rejected even though it looks rightThe gateway key and a provider key (for "proxy mode") got swappedOnly the Inference.net project key goes in INFERENCE_API_KEY; a provider key (e.g. an OpenAI key) is only for the separate x-inference-provider-api-key proxy-mode header, not used by this catalog entry

See also​