Skip to main content

Vispark Provider Guide

Vispark is a Tier-2 catalog provider: its integration is one JSON file (src/lib/providers/catalog/vispark.json) rather than hand-written code. The catalog values on this page come from that file.

Verification status: this entry is docs- and roster-verified, not yet live-verified. The fields below come from the text of Vispark's public site (https://lab.vispark.in/) and unauthenticated GET requests to https://api.lab.vispark.in/v1 — no account was created and no API key was used to build it. evidence.liveMatrix is null until someone runs the live capability matrix with a real key (see Verification status below).

Key Facts​

  • Provider id: vispark
  • Protocol: OpenAI-compatible (/chat/completions) — the site's OpenAI Compatibility panel says "This model is compatible with OpenAI-compatible clients."
  • Base URL: https://api.lab.vispark.in/v1 — that panel gives the Base URL https://api.lab.vispark.in. Unauthenticated GETs made 2026-09-29: https://api.lab.vispark.in/v1/models and https://api.lab.vispark.in/models returned the same 3-model body; https://api.lab.vispark.in/v1/chat/completions and https://api.lab.vispark.in/chat/completions both answered HTTP 405
  • Default model: vispark/vision-large (NeuroLink default)
  • Models in catalog: 3 — the roster returned 3 models on 2026-09-29
  • Streaming: true — the Vision entry on the site reads "Supports up to 1 million tokens per request with real-time streaming capabilities."
  • Tool calling: true — the Vision entry reads "Supports Tool Calling and Continual Learning; updates weights in real-time.", and the roster lists tools in supported_features for the three models. No POST request was sent
  • Tools while streaming: false in the catalog — the site's "Tool Calling Guide" has a "Streaming with Tool Calls" section whose sample chunk shows "type": "tool_calls" and "finished": true. No POST request was sent
  • Structured output: false in the catalog — the roster lists json_mode and structured_outputs in supported_features. No POST request was sent
  • Structured output + tools together: false in the catalog. No POST request was sent
  • Embeddings: false in the catalog — the roster's output_modalities for the three models is text
  • Thinking: false in the catalog — the roster lists reasoning in supported_features. No POST request was sent
  • Billing policy field: no-free-tier is a schema placeholder (the schema has no unknown value), not a vendor statement; the unit strings the site shows are quoted under Billing
  • Key format: apiKeyFormat is null in the catalog

Quick Start​

1. Get an API key​

  1. Visit: https://lab.vispark.in/profile and sign in — the playground text reads "Sign in to run this model.", and the site links the Vispark account system at https://account.vispark.in
  2. Open API Keys in the profile sidebar (the panel is titled "API Keys" with the subtitle "Manage your API keys"), choose "Add Key" and complete the "Create New API Key" dialog
  3. Authentication is HTTP Bearer: the OpenAI Compatibility panel says "Use your standard API Key as a Bearer Token." and shows Authorization: Bearer <your_api_key>. The site's developer-guide text for the Vision model also lists "Endpoint: /model/text/vision" and "Authentication: X-API-Key: your_api_key_here", and the ephemeral-key panel "Using on Model Endpoints" shows an X-API-Key header and lists "POST /chat/completions" under "Supported endpoints:". Neither header scheme was exercised: no POST request was sent
  4. Billing as the site shows it: the Account tab of the profile page has a "Buy More Units" panel with the strings "₹1 (INR) = 1 unit" and "$1 (USD) = ~95 units"
  5. Set VISPARK_API_KEY in your .env file

2. Configure​

export VISPARK_API_KEY=your-api-key
export VISPARK_MODEL=vispark/vision-large   # optional — overrides the default model
export VISPARK_BASE_URL=https://api.lab.vispark.in/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: "vispark",
  model: "vispark/vision-large",
});

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

Per-request credentials are passed like this:

await neurolink.generate({
  input: { text: "Hello" },
  provider: "vispark",
  credentials: { vispark: { apiKey: process.env.VISPARK_API_KEY } },
});

Billing​

The Account tab of https://lab.vispark.in/profile has a "Buy More Units" panel with the strings "₹1 (INR) = 1 unit" and "$1 (USD) = 95 units", and the Insufficient Units dialog carries the text "$0.01 = ₹1 = 1 unit." The Vision model's price text is quoted in Models. The no-free-tier value in the catalog is a schema placeholder, not a vendor statement.

Models​

ModelContextMax outputVision$/M in · outNotes
vispark/vision-large ⭐1,000,00065,536yes$7.37 / $22.11NeuroLink default. Roster description: "Vision Large: India's most advanced foundational intelligence." The site's size entry reads "Slowest processing but highest intelligence and accuracy".
vispark/vision-medium1,000,00065,536yes$4.21 / $12.63Fallback. Roster description: "Vision Medium: A versatile foundational model built for the world on Indian infrastructure." The site's size entry reads "Balanced processing speed with good intelligence".
vispark/vision-small1,000,00065,536yes$1.05 / $3.16Fallback. Roster description: "Vision Small: India's fast and efficient foundational AI." The site's size entry reads "Fastest processing speed with basic intelligence".

Context, max output, prices and modalities come from an unauthenticated GET https://api.lab.vispark.in/v1/models call made 2026-09-29 (3 models; 3 in this catalog): context_length 1000000, max_output_length 65536, and input_modalities text, image, audio, video, file for the three models. The $/M figures are the roster's pricing.prompt and pricing.completion strings times 1,000,000 ("0.00000737" and "0.00002211" for vispark/vision-large, "0.00000421" and "0.00001263" for vispark/vision-medium, "0.00000105" and "0.00000316" for vispark/vision-small).

The site's price text for the Vision sizes, from the script https://lab.vispark.in/assets/index-WmfFPhvJ.js that https://lab.vispark.in/ loads (retrieved 2026-09-29):

SizePrice text as the site shows it
small"₹10 input/₹30 output per 100k tokens. Minimum 600 input tokens charged."
medium"₹40 input/₹120 output per 100k tokens. Minimum 600 input tokens charged."
large"₹70 input/₹210 output per 100k tokens. Minimum 600 input tokens charged."

For vispark/vision-small, ₹10 per 100k input tokens is ₹100 per 1M, which at "$1 (USD) = ~95 units" is about $1.05, the roster's figure.

Vision: the three models are vision: true. The roster lists image in input_modalities, and the site's playground text reads "Supports images, audio, video, and PDFs (max 25MB each)". The Vision model's Limitations text reads "Combined content: 980,000 tokens maximum".

Each model's status in the JSON is production, a placeholder for a value the catalog schema requires; it is not a vendor statement.

models.defaultContextWindow (128,000) and models.defaultMaxOutputTokens (16,384) are NeuroLink placeholders for model ids outside the catalog, not vendor figures.

Fallback order when the default is unavailable: vispark/vision-medium → vispark/vision-small. The runtime fallback model name the loader derives (fallbacks[1]) is vispark/vision-small.

Vendor-stated error handling​

These lines are quoted from the Vision model's Limitations text on the site (script https://lab.vispark.in/assets/index-WmfFPhvJ.js, retrieved 2026-09-29). They are the vendor's wording, not NeuroLink behaviour, and the catalog's errorRules (401, 402, 429) carry them:

  • "Check HTTP status codes (200, 400, 401, 402, 429)"
  • "Handle rate limiting with exponential backoff"
  • "Validate API keys and sufficient units"
  • "API rate limit: 240 requests/minute"

The site's script also uses "Insufficient units for this operation" as the fallback message for HTTP 402.

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 Vispark — docs- and roster-verified, not live-verified:

ProbeResult
Rosterunauthenticated GET /v1/models, HTTP 200, 3 models, 2026-09-29 — no key used; GET https://api.lab.vispark.in/models returned the same body
Billingpublic script strings "₹1 (INR) = 1 unit" and "$1 (USD) = ~95 units" from the Buy More Units panel, 2026-09-29
Auth-failure shapeunauthenticated GET https://api.lab.vispark.in/v1/chat/completions returned HTTP 405 with an HTML body titled "405 Method Not Allowed"; no authProbe is recorded, and errorRules carry the status codes and sentences of the site's Limitations text
Tools / structured outputthe roster lists tools, json_mode and structured_outputs in supported_features, and the Vision entry reads "Supports Tool Calling and Continual Learning; updates weights in real-time."; none was exercised, 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=vispark 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.

See also​