API

Use the data

The dataset is free to use under CC BY 4.0. Fetch it at build time or at startup and size your requests to fit.

Endpoints

• /v1/capabilities.json — every pair, every limit, with evidence.
• /v1/capabilities.csv — one row per pair, for spreadsheets.
• /v1/changes.json — every recorded change, by date.

All endpoints are static files with CORS enabled. The schema is versioned; breaking changes get a new /v2 path.

Shape

Each entry is one model on one provider. Each limit has a status:

• measured — value is the largest passing request, fails_at the first failure.
• stated — the provider's error states the limit (value).
• at_least — no failure up to our probe's cap; the limit is higher than value.
• inconclusive — the search stopped early; at_least, if present, passed.
• unsupported / not_tested.

Bytes are in bytes; counts are counts. An excerpt:

{
  "id": "fireworks/kimi-k3",
  "provider": {
    "id": "fireworks",
    "name": "Fireworks",
    "via": null
  },
  "model": {
    "id": "moonshotai/kimi-k3",
    "name": "Kimi K3",
    "vendor": "MoonshotAI",
    "api_id": "accounts/fireworks/models/kimi-k3",
    "input_modalities": [
      "text",
      "image",
      "video"
    ]
  },
  "measured_at": "2026-10-31",
  "limits": {
    "images_per_request": {
      "status": "measured",
      "value": 120,
      "fails_at": 121,
      "failure": "image_count",
      "stated": 120,
      "evidence": "Too many images were provided, we currently limit the number of images per conversation to 120"
    },
    "video_file_bytes": {
      "status": "measured",
      "value": 96000000,
      "fails_at": 128000000,
      "failure": "other",
      "evidence": "Video <url> download size exceeded maximum size limit of 100.0 MB"
    }
  },
  "accepts": {
    "image_inline": true,
    "image_url": true,
    "video": "yes",
    "request_fields": {
      "stream_options": true,
      "usage": false,
      "reasoning": false,
      "response_format": true,
      "tools": true
    }
  },
  "behaviour": {
    "strict_tool_message_order": false
  },
  "notes": [
    "Fireworks counts every image in the messages array (\"per conversation\"), not per message."
  ],
  "featured": true
}

Example: fit frames to a provider

const data = await fetch("https://vlm.fyi/v1/capabilities.json").then((r) => r.json());
const entry = data.entries.find((e) => e.id === "fireworks/kimi-k3");
const cap = entry.limits.frames_768x432;
// Leave headroom: limits change, and other messages use context too.
const maxFrames = Math.floor((cap.value ?? 0) * 0.85);