Fusion model

Not yet released

See More

Fusion is not in any released litellm version or on litellm main yet. It lives on the litellm_fusion_router branch and ships once litellm PR #39224 merges. Until then, the snippets on this page fail with BadRequestError: LLM Provider NOT provided ... You passed model=litellm/fusion-1

litellm/fusion-1 is a LiteLLM-provided virtual model for requests that benefit from independent model perspectives. It sends the request to a configurable panel in parallel, asks a judge to compare the successful responses, then has the judge synthesize one public response.

The request and response keep the native shape of the SDK method you call. Fusion works with Chat Completions, Anthropic Messages, and Responses.

The judge first produces a structured comparison covering consensus, contradictions, coverage gaps, unique insights, and blind spots. LiteLLM feeds that private analysis back to the judge, which writes the final answer. The public response.model identifies the concrete judge model that produced it.

Quick start

Set credentials for every provider used by the panel and judge. The same fusion configuration works across all three SDK interfaces.

Chat Completions

import litellm

fusion = {
    "models": [
        "openai/gpt-4o-mini",
        "anthropic/claude-haiku-4-5",
    ],
    "judge": {
        "model": "openai/gpt-4o-mini",
        "criteria": "Prioritize correctness and verifiable evidence.",
    },
}

response = litellm.completion(
    model="litellm/fusion-1",
    messages=[
        {
            "role": "user",
            "content": "Review this database migration plan for operational risks.",
        }
    ],
    fusion=fusion,
)

print(response.model)
print(response.choices[0].message.content)

Anthropic Messages

import litellm

fusion = {
    "models": [
        "openai/gpt-4o-mini",
        "anthropic/claude-haiku-4-5",
    ],
    "judge": {
        "model": "openai/gpt-4o-mini",
        "criteria": "Prioritize correctness and verifiable evidence.",
    },
}

response = litellm.anthropic.messages.create(
    model="litellm/fusion-1",
    max_tokens=2048,
    messages=[
        {
            "role": "user",
            "content": "Review this database migration plan for operational risks.",
        }
    ],
    fusion=fusion,
)

print(response["model"])
print(response["content"][0]["text"])

Responses

import litellm

fusion = {
    "models": [
        "openai/gpt-4o-mini",
        "anthropic/claude-haiku-4-5",
    ],
    "judge": {
        "model": "openai/gpt-4o-mini",
        "criteria": "Prioritize correctness and verifiable evidence.",
    },
}

response = litellm.responses(
    model="litellm/fusion-1",
    input="Review this database migration plan for operational risks.",
    fusion=fusion,
)

print(response.model)
print(response.output_text)

Calling litellm/fusion-1 always runs deliberation. There is no separate flag to enable it. You can omit fusion to use the default panel and judge.

Response behavior

Fusion returns one normal response in the format of the SDK method you called. It does not return separate panel responses or the private judge comparison. The response's hidden Fusion metadata reports whether deliberation ran and how many panel calls succeeded or failed, without including their content. Its hidden router field identifies litellm/fusion-1.

Configuration

Pass a fusion dictionary alongside model="litellm/fusion-1". Every field is optional.

Field	Default	Description
models	LiteLLM default panel	One to eight models that answer independently in parallel.
judge.model	LiteLLM default judge	Model that compares successful panel responses and writes the final answer.
judge.criteria	None	Instructions used to compare the panel responses.
max_completion_tokens	16000	Maximum output tokens, including reasoning, for each internal call.
reasoning	Provider default	Reasoning effort forwarded to panel and judge calls.
temperature	Provider default	Temperature forwarded to panel calls. The judge uses temperature 0.

Tools and streaming

Standard tools remain available on the public request. Client tool schemas are kept private from panel and comparison calls. The judge receives them when it authors the final response.

Streaming uses the async SDK methods: litellm.acompletion(), litellm.anthropic.messages.acreate(), or litellm.aresponses(). Synchronous streaming is rejected with guidance to the matching async method.

Cost and recursion

Fusion makes several provider calls for one public request. Cost and latency increase with panel size and the configured models. Panel and judge models cannot recursively invoke litellm/fusion-1; deliberation is bounded to one level.