Introduction
Ling 3.1 Flash is a hybrid-reasoning mixture-of-experts model from inclusionAI, released on 29 September 2026. The publisher describes it as 560B parameters in total with 25B active per token. Here it is one OpenAI-compatible chat endpoint: up to 262,144 tokens of text in, up to 262,144 tokens out, tools and streaming on every request.
| Field | Value |
|---|---|
| Input | Text |
| Output | Text |
| Input budget | 262,144 tokens |
| max_tokens ceiling | 262,144 |
| Inputs | text |
Key features & highlights
A very large model that only runs a slice of itself
inclusionAI's model card describes Ling 3.1 Flash as a hybrid-linear mixture-of-experts: 560B parameters in total, with eight of 512 routed experts plus one shared expert active for each token, about 25B parameters. That design is what the name Flash points at: a large model that is meant to answer at the pace of a much smaller one.
Think when it helps, skip it when it does not
reasoning_effort takes none, low, medium or high, and medium is the default. Use none or low for classification, extraction and short chat turns where latency matters, and move up to high for a multi-step problem. The publisher calls the model a native hybrid-reasoning model, so one endpoint covers both styles.
A quarter of a million tokens in one request
262,144 tokens of input is roughly 190,000 English words, enough for a long contract, a support thread with its attachments pasted as text, or a mid-sized module of code. The model card lists a 1M-token window for the open model; this endpoint accepts 262,144, and that is the figure to plan around.
Tool calls and streaming in the same call
Send up to 128 tool definitions, let the model return a function call, run it in your application and send the result back as a tool message. Replies stream as they are written, and response_format lets you ask for a JSON object when a downstream service expects one.
Request parameters
| Parameter | Required | Description |
|---|---|---|
messages | Yes | Send Ling 3.1 Flash an ordered conversation containing the source material for product and delivery-date extraction; keep tool results paired with |
max_tokens | No | For Ling 3.1 Flash, |
reasoning_effort | No | Ling 3.1 Flash accepts none, low, medium, high, with medium as the default; compare effort settings on product and delivery-date extraction while monitoring output-token usage. |
stop | No | Stop Ling 3.1 Flash with one sequence or up to four sequences; the matched marker is omitted from the final text. at least 1 characters · at least 1 items · at most 4 items |
temperature | No | Use temperature to adjust sampling in Ling 3.1 Flash; keep its default while evaluating product and delivery-date extraction unless you need a different answer style. |
top_p | No |
|
tools | No | Give Ling 3.1 Flash up to 128 OpenAI-format function definitions for product and delivery-date extraction; your application checks arguments and performs the proposed operations. at most 128 items |
parallel_tool_calls | No | In a Ling 3.1 Flash turn, this switch permits proposals for several independent functions; the application still controls their execution. Default: |
tool_choice | No | For Ling 3.1 Flash, |
response_format | No | Request text or structured JSON from Ling 3.1 Flash for product and delivery-date extraction, then check the JSON against the fields your application expects. |
Pricing
Input $0.075 / Output $0.225 per 1 million tokens
Use cases
High-volume classification and extraction
Tag support tickets, pull names, dates and amounts out of emails, or route messages to the right queue. Set reasoning_effort to none or low, keep temperature low, and ask for a JSON object so each answer can be parsed without cleanup.
Fast chat for products and assistants
Put the persona and house rules in the system message, send the conversation oldest first, and stream the reply to the user. The large context means long histories can be sent as written instead of being summarized.
Agent loops with tools
Give the model a search function, a database lookup or a ticketing action and let it decide when to call them. The publisher lists terminal-based tool use and software-engineering work among the model's intended uses.
Reading long documents
Paste a policy manual, a transcript or a code module, put the material first and the question last, and ask for the passage that supports each claim. Raise reasoning_effort when the answer depends on putting several sections together.
Prompt guide & pro tips
Start at medium, then move one step
medium is the default. Drop to low or none for short, well-defined jobs and watch latency fall; raise to high only for problems where a wrong answer costs more than the extra tokens. Compare outputs on a few of your own examples before you change a production default.
Leave room in
max_tokensfor thinkingThe default
max_tokensis 4,096. Thinking is counted in the same budget, so raise it for long answers or at higher effort levels, or the budget can run out before the visible reply is finished. Check the finish reason before treating a reply as complete.Put long material first and the question last
Label each source, place the documents early in the conversation and ask your question at the end. Keep instructions in the system message so they are not confused with quoted text.
Return tool results with the matching id
When the model asks for a function call, run it, then send the output as a tool message that carries the same
tool_call_id. Keep the assistant's earlier tool calls in the history so the conversation stays consistent.
Ling 3.1 Flash for text extraction
Start with a customer update and ask for the product name, delivery date and the exact phrase supporting each field. This endpoint is text only: include the words from a document rather than attaching a picture of it. Its exposed input limit is 262,144 tokens, even though inclusionAI describes a 1M context for the underlying model.
Choose the reasoning budget
The default effort is medium; none, low and high are also available. Use none for a straightforward extraction and compare high on a case that needs several deductions. The default output cap is 4,096 tokens; raising it permits a longer answer without changing the token rate.
From extraction to an action
Ask for response_format JSON when a program needs the fields. Validate the values against the original update before using them. Ling can propose parallel tool calls, but your application executes the functions and returns their results; a generated delivery date alone is not a verified shipment status.
Content decisions come from the selected model. Ling 3.1 Flash: NSFW behavior can vary, and no unrestricted-content guarantee is made.
API quick start
curl --fail-with-body -sN https://api.spicyapi.ai/v1/chat/completions \
-H "Authorization: Bearer $SPICY_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: $(uuidgen)" \
-d '{
"model": "inclusionai/ling-3.1-flash/chat",
"messages": [
{
"role": "user",
"content": "Summarise the following release notes into three bullet points."
}
],
"stream": true
}' \
| jq --unbuffered -Rrj 'select(startswith("data: ")) | .[6:] | select(. != "[DONE]") | fromjson | if .error then "\n[error] \(.error.message)\n" else (.choices[0].delta.content? // empty) end'
See the API reference for authentication, the complete schema, polling, and error handling.