by Unity Nodes
docs.near.aicloud › guides › fusion

NEAR AI Cloud: Fusion

Updated 02-11-2026

View on docs.near.ai

Use server-side multi-model deliberation from OpenAI-compatible clients

Fusion runs a private, server-side deliberation across multiple models and returns one OpenAI-compatible chat completion. Your client sends one request to NEAR AI Cloud; the inference proxy fans out to the configured panel models, optionally asks a judge model for structured guidance, then synthesizes the final answer with the original model.

Fusion is available on /v1/chat/completions through the Cloud API gateway. Cloud API remains a pass-through: it bills the single request from the final aggregate usage, which includes panel, judge, and synthesis model tokens.

Request Shapes

NEAR AI Cloud accepts OpenRouter-style Fusion configuration in either a server tool or a plugin entry.

Server tool

curl https://cloud-api.near.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "Qwen/Qwen3.6-27B-FP8",
    "messages": [
      {
        "role": "user",
        "content": "Compare SQLite and Postgres for a mobile sync backend. Be concise."
      }
    ],
    "tools": [
      {
        "type": "openrouter:fusion",
        "parameters": {
          "analysis_models": [
            "Qwen/Qwen3.6-27B-FP8",
            "zai-org/GLM-5.2-FP8"
          ],
          "model": "zai-org/GLM-5.2-FP8",
          "max_completion_tokens": 512,
          "temperature": 0.2
        }
      }
    ],
    "tool_choice": "required"
  }'

Plugin

curl https://cloud-api.near.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "Qwen/Qwen3.6-27B-FP8",
    "messages": [
      {
        "role": "user",
        "content": "Compare SQLite and Postgres for a mobile sync backend. Be concise."
      }
    ],
    "plugins": [
      {
        "id": "fusion",
        "analysis_models": [
          "Qwen/Qwen3.6-27B-FP8",
          "zai-org/GLM-5.2-FP8"
        ],
        "model": "zai-org/GLM-5.2-FP8",
        "max_completion_tokens": 512,
        "temperature": 0.2
      }
    ],
    "tool_choice": "required"
  }'

Use tool_choice: "required" when you want Fusion to run immediately. If you omit it, the outer model can decide whether to call Fusion.

[!NOTE]

Use the top-level request model for the model that should synthesize the final answer. The Fusion parameters.model or plugin model field selects the judge model. NEAR AI Cloud does not route the OpenRouter virtual model alias openrouter/fusion; the Cloud API gateway routes by real model name before the request reaches the inference proxy.

Parameters

Field Required Description
analysis_models Yes* Panel models to ask before synthesis. Required unless the deployment configures FUSION_DEFAULT_ANALYSIS_MODELS. Model names may include or omit a leading ~; NEAR AI ignores the prefix when resolving the model.
model No Judge model. Defaults to the request model when omitted.
max_tool_calls No Maximum inner web-search tool calls for panel and judge requests.
max_completion_tokens No Token cap for panel and judge completions.
temperature No Sampling temperature for panel and judge calls.
reasoning No Reasoning configuration forwarded to panel and judge models when supported.

Response Metadata

The response remains OpenAI-compatible and may include a top-level nearai_fusion object:

{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 1200,
    "completion_tokens": 800,
    "total_tokens": 2000
  },
  "nearai_fusion": {
    "status": "invoked",
    "forced": true,
    "panel": [
      {
        "model": "Qwen/Qwen3.6-27B-FP8",
        "status": "ok",
        "verifiable": true,
        "usage": {
          "prompt_tokens": 300,
          "completion_tokens": 200,
          "total_tokens": 500
        }
      }
    ],
    "judge": {
      "model": "zai-org/GLM-5.2-FP8",
      "status": "ok"
    },
    "aggregate_usage": {
      "prompt_tokens": 1200,
      "completion_tokens": 800,
      "total_tokens": 2000
    }
  }
}

Raw panel answers are not returned by default. Panel failures are reported in metadata when at least one panel succeeds; if every panel fails, the request fails with a Fusion error.

Fusion can use NEAR AI's web_context_search tool inside panel and judge calls. Include {"type": "web_context_search"} in the same tools array and set max_tool_calls in the Fusion configuration to bound inner search calls.

{
  "tools": [
    {
      "type": "openrouter:fusion",
      "parameters": {
        "analysis_models": ["Qwen/Qwen3.6-27B-FP8", "zai-org/GLM-5.2-FP8"],
        "max_tool_calls": 1
      }
    },
    { "type": "web_context_search" }
  ],
  "tool_choice": "required"
}

See Also

Back to Documentation