Skip to main content

Vultr Inference Provider Guide

Vultr Inference is a Tier-2 catalog provider: it is served by NeuroLink's OpenAI-compatible chat-completions path, so its entire integration is one JSON file (src/lib/providers/catalog/vultr-inference.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. The model data, endpoint and error statuses below come from Vultr's public pages and unauthenticated GET requests, and no account was created and no API key was used to build it. billingPolicy, each model's status and the two models.default* values are required by the schema and are not Vultr statements. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

Key Facts​

  • Provider id: vultr-inference
  • Protocol: /chat/completions with bearer-token authentication. The API reference at https://api.vultrinference.com/ says "Responses follow OpenRouter's chat completion shape whichever engine serves the model", and the FAQ at https://docs.vultr.com/products/compute/serverless-inference/faq says to "replace your current inference API URL (such as OpenAI's base API URL) with Vultr’s API URL"
  • Base URL: https://api.vultrinference.com/v1
  • Default model: glm-5.2
  • Models in catalog: 12 (curated from a 24-model roster: 15 entries with a text output, 3 rerank, 3 decision, 1 embeddings, 1 transcription and 1 image). nemotron-3.5-content-safety (roster: "multimodal, multilingual content moderation"), nemotron-3-nano-omni-30b-a3b-reasoning and qwen3.8-27b are the three text-output roster entries that are not catalogued
  • Message shape: the chat-completions request message schema lists role (system, assistant, user, tool, developer) and content (type string, "The contents of the message."); the tool call id description says "Send it back unchanged as the tool_call_id of the tool result." Not exercised live
  • Streaming: supported — the chat-completions reference documents stream ("Indicates whether the response should be streamed.") and a text/event-stream response, and the roster lists streaming: true on the 12 catalogued models
  • Tool calling: model-dependent — the reference documents tools ("A list of function tools the model may call.") and tool_calls in responses; the roster lists tools among the supported parameters of the 12 catalogued models, and the nemotron-3.5-content-safety entry does not list it
  • Tools while streaming: declared (true) — the streaming payload schema in the OpenAPI document embedded in https://api.vultrinference.com/ says "A delta carries only the fields it adds (role, content, reasoning, tool_calls)". Not exercised live
  • Structured output: not declared (false) — the chat-completions request schema lists model, messages, continue_final_message, stream, max_tokens, max_completion_tokens, reasoning_effort, reasoning, n, seed, temperature, top_p, frequency_penalty, presence_penalty, stop, logprobs, top_logprobs, tool_choice and tools, and no response_format
  • Structured output + tools together: not declared (false) — no combined probe was run without credentials
  • Embeddings: not declared on this catalog entry. The reference lists a POST /embeddings route and the roster lists qwen3-embedding-4b, but NeuroLink's generic ConfiguredOpenAICompatProvider (which Tier-2 catalog entries use) does not implement embed()/embedMany(), so this flag tracks that, not the vendor's own API surface
  • Thinking: not declared (false) — the reference documents reasoning_effort and a reasoning request object, and this entry has not been run with NeuroLink's thinkingLevel option
  • Billing: the provisioning page lists a charges note to acknowledge, and the FAQ describes a Usage tab (see Quick Start). The entry records no-free-tier because the schema has no unknown value; it is not a Vultr statement
  • Key format: none declared

Quick Start​

1. Get an API key​

  1. Visit: https://docs.vultr.com/products/compute/serverless-inference/provisioning and follow the Vultr Console steps: Products, Serverless, Inference, Add Serverless Inference, then provide a Label, "acknowledge the list of supported models and the charges note", and click Add Serverless Inference
  2. Open the new subscription's management page and copy the API key from its overview page (https://docs.vultr.com/products/compute/serverless-inference/management/connection: "API key is necessary to send any request to the inference endpoints")
  3. Charges: the provisioning page above lists the charges note among the items to acknowledge, and the FAQ (https://docs.vultr.com/products/compute/serverless-inference/faq) says the subscription's Usage tab shows "details on your current token usage, overage, and any associated costs"
  4. Set VULTR_INFERENCE_API_KEY in your .env file (the authentication example on https://api.vultrinference.com/ writes ${VULTR_API_KEY}, which NeuroLink does not read)

The FAQ also says the key can be regenerated from the Overview page in the Vultr Console and that "This will invalidate the previous API key".

2. Configure​

export VULTR_INFERENCE_API_KEY=your-api-key
export VULTR_INFERENCE_MODEL=glm-5.2   # optional — overrides the default model
export VULTR_INFERENCE_BASE_URL=https://api.vultrinference.com/v1   # optional — proxy or self-hosted gateway

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: "vultr-inference",
  model: "glm-5.2",
});

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

Per-request credentials:

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

Models​

ModelContextMax outputVision$/M in · outNotes
glm-5.2 ⭐1,048,5761,048,576yes$0.75 / $3NeuroLink default; roster: "GLM-5.2 is Z.ai's flagship model for long-horizon tasks".
deepseek-v4.1-flash1,048,5761,048,576yes$0.15 / $0.6Fallback; roster: "support for reasoning and tool calling".
qwen3.8-flash-next262,144262,144yes$0.1 / $0.2Fallback; roster: "reasoning, and tool calling".
mimo-v2.6-flash-rl1,048,5761,048,576yes$0.1 / $0.25Fallback; roster: "Supports reasoning and tool calling, served at FP8."
deepseek-v4-flash-07311,048,5761,048,576yes$0.1 / $0.25Roster: "DeepSeek's official V4 Flash release".
glm-5.31,048,5761,048,576yes$0.75 / $3Roster: "Served at NVFP4 across a 1M-token context."
glm-5.3-flash1,048,5761,048,576yes$0.1 / $0.35Roster: "Z.ai's first natively multimodal GLM-5 model".
glm-5.x-menthol202,752202,752yes$0.4 / $1.75Roster name "GLM 5"; "Z.ai's GLM-5, a 744B-parameter MoE flagship for agentic reasoning and coding".
laguna-s-2.11,048,5761,048,576yes$0.09 / $0.18Roster: "designed for agentic coding and long-horizon work".
mimo-v2.6-pro-rl1,048,5761,048,576yes$0.4 / $0.8Roster: "Supports reasoning and tool calling, served at FP8."
minimax-m3524,288524,288yes$0.2 / $0.9The roster description text says a 1M-token context; the max_context_length field says 524,288.
muse-glimmer-30b131,072131,072yes$0.25 / $1Roster: "a dense multimodal model for reasoning, coding, and tool use".

Context, max output and prices come from the unauthenticated GET https://api.vultrinference.com/v1/models call made 2026-09-29 (24 entries; 12 curated into this catalog). Context is the text input max_context_length, max output is the output modality's max_length, and prices are the roster's cost_usd per token for the prompt and completion entries, multiplied by one million.

models.defaultContextWindow (131,072) and models.defaultMaxOutputTokens (32,768) are placeholders for ids outside the catalog that the vendor does not publish: 131,072 is the context the roster lists for muse-glimmer-30b, and 32,768 is the default the chat-completions page shows for max_completion_tokens. Each model's status is production; the field is required by the schema and the value is not a Vultr statement.

Vision: the roster lists image among the input modalities of the 12 catalogued models (sources url and base64; formats image/png, image/jpeg, image/webp, image/gif), hence vision: true. See Message shape under Key Facts for the request message schema. No image request was sent.

Fallback order when the default is unavailable: deepseek-v4.1-flash → qwen3.8-flash-next → mimo-v2.6-flash-rl. The runtime fallback model name the loader derives (fallbacks[1]) is qwen3.8-flash-next.

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 currently records for Vultr Inference — docs- and roster-verified, not live-verified:

ProbeResult
Rosterunauthenticated GET /v1/models, HTTP 200, 24 models, 2026-09-29 — no key used
Auth-failure shapeunauthenticated GET /v1/models/all returned 401 {"message":"No API key provided"} (the same body on GET /v1/usage and GET /v1/models/glm-5.2); the body carries no error code. The chat-completions reference lists 401 "Unauthorized" and 429 "Rate limit exceeded". GET /v1/chat/completions answered 404 and no POST was sent, so errorRules match the statuses 401 and 429
Billingthe provisioning page's Console steps say "acknowledge the list of supported models and the charges note"; the FAQ's Usage tab shows "details on your current token usage, overage, and any associated costs" — billingPolicy is the schema value no-free-tier, which is not a Vultr statement
Tools / structured outputtools and tool_calls are documented in the chat-completions reference; response_format is not in its request schema. Neither was exercised live, and no combined tools+schema request was sent — structuredOutputWithTools stays false
Live capability sweepnot run. evidence.liveMatrix is null. Before treating this provider as production-ready, run npx tsx test/continuous-test-suite-provider-matrix.ts --provider=vultr-inference with a real key and record the result.

Do not treat this entry as equivalent to a live-verified Tier-2 provider (e.g. FriendliAI, Novita AI) until that live matrix has been run and evidence.liveMatrix is filled in.

Documented error statuses​

The chat-completions reference at https://api.vultrinference.com/ lists these responses for POST /chat/completions, and says errors use OpenRouter's envelope, {"error": {"code", "message", "type", "metadata": {"error_type"}}}:

StatusVultr's description
400"Bad Request, including context_length_exceeded"
401"Unauthorized"
422"Validation Error"; the reasoning_effort description says "unsupported levels return 422"
429"Rate limit exceeded"
502"The model could not be reached"
504"The model timed out"

See also​