> This page is for Model APIs. > Sarvam Voice Agents (Samvaad) is a voice AI agent platform for Indian languages: build, test and deploy AI calling agents on your own phone numbers, from ₹3 a minute. Start here: https://docs.sarvam.ai/voice-agents/why-voice-agents. > For clean Markdown of any page, append `.md` to the page URL. > For a complete documentation index, see https://docs.sarvam.ai/llms.txt. For a section index, append `/llms.txt` to any section URL (for example https://docs.sarvam.ai/voice-agents/llms.txt). > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.sarvam.ai/_mcp/server. # Responses API > Create stateless text, reasoning, structured-output, and tool-calling responses with Sarvam's OpenAI-compatible Responses API. The Responses API provides an OpenAI-compatible interface for text generation, reasoning, structured output, and tool calling. **Streaming is optional**: the same pattern as [Chat Completion V2](/api-reference/chat/chat-completions-v2): omit `stream` or set it to `false` for one JSON response (default). > **Note** > > The Responses API supports `sarvam-105b`, `glm5.3`, `glm5.3-flash`, and `gemma4`, > plus the [beta models](/api-reference/beta-apis) `deepseekv4.1-flash-1017`, `deepseekv4-pro`, > `kimi-k3`, `mimo-v2.6-flash`, and `mimo-v2.6-pro`. > **Tip** > > Moving Responses API code from [OpenRouter](/api/token-factory/migrations/from-openrouter/responses-api) or [Fireworks AI](/api/token-factory/migrations/from-fireworks/responses-api)? See the Responses API migration guides. ## Create a response Use [`POST /v2/responses`](/api-reference/responses/create). The endpoint is **stateless**: send the complete conversation in `input` on every request. Omit `store` or set it to `false` (`true` returns **`400`**). Nothing is stored for later retrieval, `GET /v2/responses/{id}` returns **`404`**. `GET /v2/responses` and `DELETE /v2/responses/{id}` return **`405`** (documented in the [API Reference](/api-reference/responses/create) alongside `POST /v2/responses`). #### Python ```python from sarvamai import SarvamAI client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY") response = client.responses.create( model="glm5.3", input="Explain database indexing in three short points.", store=False, ) print(response.output) ``` #### JavaScript ```javascript import { SarvamAIClient } from "sarvamai"; const client = new SarvamAIClient({ apiSubscriptionKey: "YOUR_SARVAM_API_KEY", }); async function main() { const response = await client.responses.create({ model: "glm5.3", input: "Explain database indexing in three short points.", store: false, }); console.log(response.output); } main(); ``` #### cURL ```bash curl -X POST https://api.sarvam.ai/v2/responses \ -H "api-subscription-key: $SARVAM_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "glm5.3", "input": "Explain database indexing in three short points.", "store": false }' ``` > **Note** > > The examples above do **not** set `stream`. That returns a single JSON body (**HTTP 200**), > same as Chat Completion V2 when `stream` is omitted. Only add `"stream": true` when you > want server-sent events. ## Stream a response (optional) Set `stream` to `true` when you want named server-sent events instead of one JSON body. Streams end with `response.completed`, `response.incomplete`, or `response.failed`; they do **not** emit a `[DONE]` sentinel. On `glm5.3`, streamed items can **nest**: pair events by `item_id`. Usage appears on the terminal event (there is no `stream_options`). #### Python ```python from sarvamai import SarvamAI client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY") stream = client.responses.create( model="glm5.3", input="Explain database indexing in three short points.", store=False, stream=True, ) for event in stream: print(event.type) ``` #### JavaScript ```javascript import { SarvamAIClient } from "sarvamai"; const client = new SarvamAIClient({ apiSubscriptionKey: "YOUR_SARVAM_API_KEY", }); async function main() { const stream = await client.responses.create({ model: "glm5.3", input: "Explain database indexing in three short points.", store: false, stream: true, }); for await (const event of stream) { console.log(event.type); } } main(); ``` #### cURL ```bash curl -N -X POST https://api.sarvam.ai/v2/responses \ -H "api-subscription-key: $SARVAM_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "glm5.3", "input": "Explain database indexing in three short points.", "store": false, "stream": true }' ``` ## Continue a conversation The endpoint does not store response state. To continue a conversation, send the prior messages and outputs again in `input`. Do not use `previous_response_id` or `conversation` , both return **`400`**. ```python from sarvamai import SarvamAI client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY") response = client.responses.create( model="gemma4", input=[ {"role": "user", "content": "Name one advantage of a B-tree index."}, {"role": "assistant", "content": "It keeps lookups efficient as data grows."}, {"role": "user", "content": "When should I avoid one?"}, ], store=False, ) ``` For message parts in `input`, text uses `{"type": "input_text", "text": "…"}` (not chat completions' `{"type": "text"}`). ## Reasoning The reasoning models (`glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3`, `mimo-v2.6-flash`, and `mimo-v2.6-pro`) support three modes, `low`, `high`, and `max`, and Chat Completion V2 accepts all three directly. `gemma4` reasons only when you set an effort. On this route (Responses), the models differ: | Model | Accepted `reasoning.effort` values on this route | | ---------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------- | | `glm5.3`, `glm5.3-flash` | `"low"` or `"high"` only; the literal string `"max"` returns `400`. Omitting the field gives `max` on `glm5.3` and `high` on `glm5.3-flash`. | | `gemma4` | `"low"` or `"high"`. No reasoning when the field is omitted. | | `deepseekv4.1-flash-1017` | `"low"`, `"high"`, or `"max"`, all three work directly, same as on chat. Default `high`. `"none"` turns thinking off. | | `deepseekv4-pro` | `"low"`, `"high"`, or `"max"`; default `high`. `"none"` turns thinking off. | | `kimi-k3` | `"low"`, `"high"`, or `"max"`; default `max`. Thinking is always on. | | `mimo-v2.6-flash`, `mimo-v2.6-pro` | Thinking is on by default. `"none"` turns it off; any other value only switches it on. | **Chat trace:** `choices[].message.reasoning_content`. **Responses trace:** reasoning models return a **`reasoning` item** in `output` when they reason. Read the trace from its **`content[0].text`**. Its `summary` is empty unless you set the optional **`reasoning.summary`** to `auto` or `detailed`. At `low` effort, `glm5.3` often skips reasoning on simple prompts and returns no reasoning item. `gemma4` reasons when you set an effort but doesn't return a reasoning item; its reasoning shows only in `usage.output_tokens_details.reasoning_tokens`. Reasoning tokens count toward **`max_output_tokens`**, which has **no default cap** on this route, omitting it means unbounded generation. On chat, the equivalent budget is **`max_tokens`**, which defaults to `2048` on `glm5.3` only. Set it explicitly. See the [Chat Completion V2 API Reference](/api-reference/chat/chat-completions-v2). ## Defaults when omitted (`glm5.3`) | Field | Default when omitted | | ------------------- | -------------------------------- | | `temperature` | `1` | | `top_p` | `0.95` | | `max_output_tokens` | No default, unbounded generation | | `stream` | `false` | ## Parameters | Parameter | Required | Description | | --------------------- | -------: | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | `model` | Yes | `sarvam-105b`, `glm5.3`, `glm5.3-flash`, `gemma4`, `deepseekv4.1-flash-1017`, `deepseekv4-pro`, `kimi-k3`, `mimo-v2.6-flash`, or `mimo-v2.6-pro`. | | `input` | Yes | A prompt string or the complete conversation as input items. | | `stream` | No | Optional, omit or `false` for one JSON body (default **`false`** on `glm5.3`); `true` for named SSE events, same pattern as chat completions. | | `max_output_tokens` | No | Cap on generated tokens including reasoning. **No default**: omitting it means unbounded generation. Always set this explicitly in production. | | `temperature` | No | `0`–`2`; default **`1`** on `glm5.3`. | | `top_p` | No | Default **`0.95`** on `glm5.3`; must be > `0`. | | `top_k` | No | Not supported on `glm5.3`; any value returns `400`. | | `reasoning` | No | Optional on reasoning models. `effort` accepts `"low"` or `"high"` on `glm5.3` (omit for `max`, the string `"max"` is `400`), or all three values directly on `deepseekv4.1-flash-1017` (omit for `high`). `summary` accepts `"auto"` or `"detailed"`. | | `text` | No | Output format in `text.format` (`text`, `json_object`, `json_schema`). | | `tools` | No | **Flat** function tools (`name` at the top level). Nested chat tool shape is **`400`**. | | `tool_choice` | No | `auto`, `none`, `required`, or a named function. | | `parallel_tool_calls` | No | **`false`** limits a turn to one tool call on every open-weight model. `sarvam-105b` ignores it on this route. | | `n` | No | Not supported on the open-weight models: any value, even `1`, returns `400`. | | `store` | No | Must be omitted or `false`. | ### Refused or unsupported on this route These return **`400`** (the error names the field): `store: true`, `background: true`, `previous_response_id`, `conversation`, `max_tool_calls`, `prompt`, `item_reference`. On Responses, the open-weight models return `400` for `logprobs` and `repetition_penalty`. For log probabilities on `glm5.3-flash` and `deepseekv4.1-flash-1017`, send `top_logprobs` with `include: ["message.output_text.logprobs"]` instead; `glm5.3` and `gemma4` don't support them on this route. On Chat Completion V2, `logprobs` and `top_logprobs` are supported on every model except `glm5.3`, which returns a clean `400`. See [Chat Completion V2](/api-reference/chat/chat-completions-v2) for details. See the [Responses API Reference](/api-reference/responses/create) for the complete request, response, error, and streaming-event schemas. Model-specific limits for `glm5.3` are also summarized on the [GLM-5.3 model page](/api/token-factory/models/glm-5-3).