Deterministic AI your contracts can settle on.

Somnia LLM Inference runs large language models inside Somnia. Every validator in a subcommittee runs the same pinned model on the same input and must return a byte-identical answer, so the contract gets an answer the validators agreed on.

Read the docs

Request 51870inferString4 of 4 identicalIllustrative

Input

system: You are a market moderator. Answer accept or reject.
prompt: Is this question clear and resolvable? "Will the venue publish its 2026 annual report by 31 March 2027?"
allowedValues: ["accept", "reject"]
chainOfThought: false

Pinned on every validator

model: open-weight
seed: fixed
temperature: 0

Same input, four validators

Validator	Model	Seed	Output	Output hash	Match
0x8f3a…21c4	pinned	fixed	accept	0x9c1e…44ab	Identical
0x2b91…7e05	pinned	fixed	accept	0x9c1e…44ab	Identical
0xc604…9a38	pinned	fixed	accept	0x9c1e…44ab	Identical
0x47de…b610	pinned	fixed	accept	0x9c1e…44ab	Identical

ConsensusByte-identical output from 4 of 4 validatorsto callbackhandleResponse("accept")

Receipts: 4 signed, one per validator, available for audit
Deposit: Unspent deposit refunded with the callback

Output modes: text from a set list, bounded numbers, chat and tool calls: 4; Output modes: text from a set list, bounded numbers, chat and tool calls; Somnia docs, LLM Inference
Validators per request, set by the calling contract: Up to 10; Validators per request, set by the calling contract; Somnia docs, Gas fees
Temperature on every validator, with a fixed seed: 0; Temperature on every validator, with a fixed seed; Somnia docs, LLM Inference
Priced per request in SOMI; see the docs for current prices: Priced per request in SOMI; see the docs for current prices

An answer every validator can repeat.

Every validator gets the same output from the same input. That is how a subcommittee agrees on one answer, and why a contract can act on it.

Get the same answer from every validator
Fixed seeds, zero temperature and a pinned open-weight model make inference deterministic, so independent validators reproduce the exact same output.
Get one answer agreed by validators
A single model endpoint can be manipulated or switched off. Answers on Somnia are agreed by a subcommittee of validators before a contract sees them.
Inspect the reasoning
Each validator signs a receipt of the steps it took, with optional chain-of-thought, so you can audit how an answer was reached.
Call the model like any other function
Contracts call the model and get the result in a callback, with no API keys and no offchain service to run.

Where the answer comes fromFrom the live page

Question	Single endpoint	Somnia
Who runs it	One provider	A validator subcommittee
Same input	Output can vary	Byte-identical output
Switched off	By the provider	Not by any one party
Audit trail	Provider logs	Signed receipts
To call it	API key and a server	A function call
Reasoning	Up to the provider	Optional, in receipts
Answer goes to	Your server	The contract callback

Choose the answer your logic needs.

Every method goes through the same agent, with the same pinned model and the same consensus.

inferString
Call inferString for single-turn text, optionally limited to a list of allowed values.
IN · allowedValues
["safe", "unsafe"]
OUT · response
"safe"
inferNumber
Call inferNumber to get an integer answer clamped to the range you set.
IN · minValue, maxValue
1 to 10
OUT · response
7
inferChat
Call inferChat for multi-turn conversations with system, user and assistant roles.
IN · messages
system, user, assistant
OUT · response
response: string
inferToolsChat
Call inferToolsChat for tool use with MCP servers, or with onchain functions that come back to your contract as calldata.
IN · tools
mcpServerUrls, onchainTools
OUT · finishReason
"stop", "tool_calls" or "max_iterations"

Put judgement into contract logic.

Create, moderate and resolve markets

Prophecy uses Somnia LLM Inference to moderate and name every new market, and Somnia Oracle uses the same deterministic stack to weigh evidence into a verdict.

Moderation limited to accept or reject
Market naming at creation
Evidence weighed into a single verdict

Somnia Oracle

New marketIllustrative

Question: "Will the venue publish its 2026 annual report by 31 March 2027?"
Moderation: inferString, allowedValues accept or reject
Result: accept, 4 of 4 identical
Name: Venue annual report, March 2027
Resolution: Somnia Oracle weighs the evidence
Verdict: One answer, with signed receipts

Send a prompt, get one agreed answer.

LLM Inference docs

moderate.tstypescript

// From docs.somnia.network/agents/base-agents/llm-inference

const abi = [{

type: 'function',

name: 'inferString',

inputs: [

{ name: 'prompt', type: 'string' },

{ name: 'system', type: 'string' },

{ name: 'chainOfThought', type: 'bool' },

{ name: 'allowedValues', type: 'string[]' }

outputs: [{ name: 'response', type: 'string' }]

}];

// Content moderation: constrain output to "safe" or "unsafe"

const calldata = encodeFunctionData({

abi,

functionName: 'inferString',

args: [

'Analyze the following text: "Check out this amazing new product!"',

'You are a content moderation assistant.',

false, // chainOfThought

['safe', 'unsafe'] // allowedValues

]

});

Returnsresponse"safe"Agreed by the subcommittee

From the LLM Inference docs. The calldata is sent through the Somnia Agents platform like any agent request; the return is illustrative.

Give the model tools, run its onchain calls.

inferToolsChat lets the model use MCP tools and propose onchain calls, which your contract runs itself.

Run MCP tools automatically

Pass MCP server URLs. The agent discovers the tools, calls them, feeds each result back to the model and continues until it is done.

Discover tools from each MCP server
Model calls a tool, the agent runs it
Result fed back to the model
finishReason "stop", response returned

MCP tool use

Run each onchain tool call from your own contract

Describe functions by Solidity signature. The model's calls come back to your contract as ABI-encoded calldata to execute before resuming.

Tool defined by signature and description
finishReason "tool_calls", calldata returned
Your contract executes each call
Resume with the results, up to maxIterations

Onchain tool use

Set the limits on every answer.

Limit answers to values you allow
inferString takes a list of allowed values, or none for free text.
Keep numbers inside your range
inferNumber clamps the answer to the minimum and maximum you set.
See how the model reasoned
Optional reasoning, recorded in the execution receipts.
Choose how answers are agreed
Accept identical answers, or receive every result and aggregate onchain.
Audit every validator's steps
Each validator signs a receipt for its run, so anyone can check it later.
Stop tool loops at a set limit
maxIterations caps model and tool round-trips for each tool-use call.

Works well with the rest of the platform.

All modules

Give your contract a model to call.

Run an inference in the Agent Explorer, or talk to the team about the judgement your contract needs.

Talk to an expert

Deterministic AI your contracts can settle on.

An answer every validator can repeat.

Get the same answer from every validator

Get one answer agreed by validators

Inspect the reasoning

Call the model like any other function

Choose the answer your logic needs.

inferString

inferNumber

inferChat

inferToolsChat

Put judgement into contract logic.

Create, moderate and resolve markets

Send a prompt, get one agreed answer.

Give the model tools, run its onchain calls.

Run MCP tools automatically

Run each onchain tool call from your own contract

Set the limits on every answer.

Limit answers to values you allow

Keep numbers inside your range

See how the model reasoned

Choose how answers are agreed

Audit every validator's steps

Stop tool loops at a set limit

Works well with the rest of the platform.

Give your contract a model to call.