# Fusion Multi-Model Reasoning

> Every model has its blind spots

Every model has its own blind spots and may miss context or favor a particular analytical perspective. Fusion lets multiple models independently analyze the same question from different perspectives, then identifies their consensus, conflicts, and omissions before generating the final answer.

## How It Works

A Fusion request consists of two stages:

![fusion](https://cdnv2-cache.udelivrs.com/2026/09/b2b5c2e34ed7e9268bf5159485637b1b_1788228486616.png)

1. **Panel:** Multiple heterogeneous models analyze the same question simultaneously and independently. Each Panel Member uses a different role prompt to produce a candidate answer from a predefined focus.
2. **Synthesis:** The Synthesizer reads the original conversation and all valid candidate answers, identifies consensus, conflicts, and omissions, and then independently generates the final answer.

Fusion does not simply vote on candidate answers or directly concatenate multiple responses. The Synthesizer still organizes the final content based on the original question and the Panel results.

## Current Model Combination

The current `ucloud/fusion` configuration uses three Panel Members and one Synthesizer:

| Stage | Model | Focus |
|---|---|---|
| Panel | `MiniMax-M3` | Core reasoning: directly analyze the question and provide an accurate, verifiable candidate answer with key reasons |
| Panel | `kimi-k3` | Completeness and boundaries: check factual completeness, boundary conditions, and potential omissions |
| Panel | `deepseek-v4-flash-0731` | Counterexamples and risks: check counterexamples, risks, and feasibility |
| Synthesis | `glm-5.3-flash` | Synthesize candidate answers, identify consensus, conflicts, and omissions, and generate the final answer |

Each request attempts to call all Panel Members in parallel. If some members fail or time out, the system can still proceed to Synthesis as long as at least one valid candidate answer is available.

> UCloud maintains the model combination based on model stability, cost, and Benchmark results, and may adjust it in the future. Refer to the Modelverse Model Logs for the models actually invoked. User customization is not currently supported.

## Fusion vs. AUTO

AUTO answers: Which model should handle this question?
Fusion takes another approach: Can several models answer the question from different perspectives at the same time, and then have their results synthesized?

| Capability | AUTO | Fusion |
|---|---|---|
| Execution | Selects one model from the candidate models | Multiple models analyze in parallel, followed by unified synthesis |
| Primary goal | Select a suitable model by balancing quality, cost, latency, and availability | Increase perspective coverage and opportunities for cross-checking on complex tasks |
| Suitable requests | General Q&A, real-time interaction, and most everyday calls | High-value, complex tasks that can tolerate longer waits |
| Cost | Close to a single-model call | More Tokens and longer latency |

## Quick Start

Fusion uses the OpenAI-compatible Chat Completions API. To integrate it seamlessly, simply replace the model name with `ucloud/fusion`.

### cURL

```bash
curl -N https://api-us-ca.umodelverse.ai/v1/chat/completions \
  -H "Authorization: Bearer ${MODELVERSE_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ucloud/fusion",
    "messages": [
      {
        "role": "user",
        "content": "Analyze which business scenarios are suitable for Fusion multi-model reasoning"
      }
    ],
    "stream": true
  }'
```

### Python

```python
from openai import OpenAI

base_url = "https://api-us-ca.umodelverse.ai/v1"
api_key = "your api key"
model = "ucloud/fusion"

client = OpenAI(
    base_url=base_url,
    api_key=api_key,
)

res = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "user",
            "content": "Analyze which business scenarios are suitable for Fusion multi-model reasoning",
        }
    ],
    stream=True,
)

for chunk in res:
    print(chunk.choices[0].delta.content or "", end="")
```

## Use Cases

Fusion is better suited to tasks that prioritize answer quality and information coverage and can tolerate longer reasoning time, such as:

- **In-depth analysis and research tasks:** Summarizing existing materials, cross-checking information, and synthesizing viewpoints.
- **Business analysis and decision support:** Comparing multiple options and identifying consensus, disagreements, potential omissions, and execution risks.
- **High-value content production:** Generating research reports, solutions, analysis reports, and other long-form content that emphasizes completeness.

Fusion is not suitable for simple Q&A, low-latency interactions, or Token-sensitive batch tasks. These requests are better served by calling a single model directly or using AUTO routing.

## Performance, Cost, and Latency

We used the DRACO Benchmark to conduct an internal comparison between the current Fusion combination and the high-performance single model `gpt-5.6-sol (reasoning max)`. Among responses that completed successfully and received valid scores, the two groups performed at similar levels, with Fusion holding a slight 0.25-point advantage:

| Metric | Fusion | `gpt-5.6-sol` |
|---|---:|---:|
| DRACO benchmark | 55.59 | 55.34 |
| Average Tokens per question | About 23,065 | About 5,229 |
| TTFT P50 | About 294 seconds | About 84 seconds |
| End-to-end latency P50 | About 349 seconds | About 86 seconds |
| Estimated invocation cost | About 1/5 of the comparison model | Baseline |

According to the `DRACO benchmark` results, because Fusion relies on parallel multi-model inference, its Token consumption, TTFT, and end-to-end latency are approximately 3–4 times those of a single model. However, its overall cost is only one-fifth that of `gpt-5.6-sol`, thanks to Modelverse's self-hosted inference capabilities, domestic model combination, and DeepSeek off-peak pricing.

Evaluation methodology:

- The scores in the table are based on responses that completed successfully and received valid scores. They do not represent the production SLA.
- Costs are estimated based on this model combination and off-peak pricing. Actual charges are subject to the Modelverse bill.
- Model upgrades, Panel combinations, prompts, and changes in upstream load may all affect the results. UCloud will continue to perform regression validation.

## Notes and Frequently Asked Questions

### How is Fusion billed?

A single Fusion call triggers multiple subrequests. You can search the Modelverse Model Logs for the models used, Tokens, and costs, and view billing details in the AstraFlow console. Because different questions produce candidate answers of different lengths, the actual cost of each Fusion request may vary.

### Does every request invoke the full Panel?

In the current version, every explicit call to `ucloud/fusion` executes the complete Panel and Synthesis process. It will gradually be integrated with AUTO routing to determine whether a request needs—and is worth—multi-model reasoning.

### Can I customize the Panel models?

Not currently. UCloud selects the model combination based on Modelverse's self-hosted inference capabilities, model stability, cost, and Benchmark results, and continuously updates and optimizes it. This prevents unverified combinations from causing unpredictable quality, cost, or stability.

### Why does Fusion respond more slowly?

Fusion must wait for multiple Panel Members to complete their independent analyses before performing Synthesis. In current internal tests, its TTFT and end-to-end latency are approximately 3–4 times those of a high-performance single model, so it is better suited to high-value tasks that can tolerate waiting.

### Can Fusion replace all single-model calls?

No. Fusion is a reasoning strategy for specific complex tasks, not the optimal default for every request. Single models or AUTO are still recommended for general Q&A, real-time interaction, and cost-sensitive batch tasks. Fusion is recommended only when multiple perspectives, cross-checking, and answer completeness are worth the additional wait.