> 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 full documentation content in one file, see https://docs.sarvam.ai/llms-full.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.sarvam.ai/_mcp/server. # Open-Source Models > Open-source models served on Sarvam infrastructure through the OpenAI-compatible chat completions API, using your existing Sarvam API key and credits. Alongside its own models, Sarvam serves a small set of **open-source** models. You reach them through the same chat completions API, with the same API key, credits, and rate limits — only the `model` field changes. **Available in beta.** GLM-5.2 and Gemma 4 31B are rolling out gradually — your API key must be [whitelisted for beta access](/api-reference/beta-apis) before these models respond. A key without access is refused with `400 invalid_request_error` before the model is called. [Contact us](/api/getting-started/help) to request access. This page focuses on the open-source models on `/v2`. The same endpoint also serves `sarvam-105b` — see [Sarvam-105B](/api/getting-started/models/sarvam-105b). `sarvam-105b-conversations` is on [`/v1`](/api/api-guides-tutorials/chat-completion/overview) only. Call [`GET /v2/models`](/api-reference/open-source/models) to see which IDs your key can use. These models are **not tuned for Indian languages**. For Indian languages, Indic scripts, or code-mixed input, use [Sarvam's own models](/api/getting-started/models) — they are trained and evaluated for it. Reach for an open-source model when you need a capability Sarvam's models do not offer, such as a very large context window or image input. ## Available models #### [GLM-5.2](/api/getting-started/models/open-source/glm-5-2) 512K token context window with tool calling and visible reasoning. Text input only. #### [Gemma 4 31B](/api/getting-started/models/open-source/gemma-4-31b) Image understanding and tool calling, answering directly without a reasoning pass. ## Capabilities at a glance | Model ID | Context window | Tool calling | Image input | Reasoning output | | -------- | -------------- | ------------ | ----------- | ------------------------------------------------------- | | `glm5.2` | 524,288 tokens | ✅ | ❌ | ✅ always on — see [Reasoning output](#reasoning-output) | | `gemma4` | 131,072 tokens | ✅ | ✅ | ❌ answers directly | For comparison, Sarvam's flagship chat model [`sarvam-105b`](/api/getting-started/models/sarvam-105b) supports tool calling with 128K context and is tuned for 23 languages (22 Indian + English). ## Using an open-source model Open-source models are served on **`/v2/chat/completions`**. The same endpoint also accepts `sarvam-105b`. Sarvam chat models — including `sarvam-105b-conversations` — are on **`/v1`**; `/v1` does not accept `glm5.2` or `gemma4`. Authentication uses the `api-subscription-key` header. A missing or invalid key is refused with **`403 invalid_api_key_error`**. [`GET /v2/models`](/api-reference/open-source/models) and [`GET /v2/models/{model_id}`](/api-reference/open-source/model) are unauthenticated and list only the models available to you. Use the [Sarvam SDK](/api/getting-started/sdks) — `client.open_source_models.chat_completions_v2(...)` in Python, or `client.openSourceModels.chatCompletionsV2(...)` in JavaScript. #### Python ```python from sarvamai import SarvamAI client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY") response = client.open_source_models.chat_completions_v2( model="glm5.2", messages=[ {"role": "user", "content": "Summarise the causes of the 2008 financial crisis."} ], temperature=0.2, max_tokens=1000, ) print(response.choices[0].message.content) ``` #### JavaScript ```javascript import { SarvamAIClient } from "sarvamai"; const client = new SarvamAIClient({ apiSubscriptionKey: "YOUR_SARVAM_API_KEY", }); async function main() { const response = await client.openSourceModels.chatCompletionsV2({ model: "glm5.2", messages: [ { role: "user", content: "Summarise the causes of the 2008 financial crisis.", }, ], temperature: 0.2, max_tokens: 1000, }); console.log(response.choices[0].message.content); } main(); ``` #### cURL ```bash curl -X POST https://api.sarvam.ai/v2/chat/completions \ -H "api-subscription-key: $SARVAM_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "messages": [ {"role": "user", "content": "Summarise the causes of the 2008 financial crisis."} ], "model": "glm5.2", "temperature": 0.2, "max_tokens": 1000 }' ``` ## Model selection In most cases the answer is a Sarvam model — reach for an open-source model only when you need a capability Sarvam's own models do not have. | If you need… | Use | Why | | -------------------------------------------------------------- | --------------------------------------------------------------- | --------------------------------------------------------------------------------------------- | | Indian languages, Indic scripts, romanised or code-mixed input | [`sarvam-105b`](/api/getting-started/models/sarvam-105b) | Trained and evaluated for it | | Highest quality reasoning or agentic work | [`sarvam-105b`](/api/getting-started/models/sarvam-105b) | Flagship model, Sarvam-published benchmarks | | Wikipedia-grounded factual answers (`wiki_grounding`) | [`sarvam-105b`](/api/getting-started/models/sarvam-105b) | **Sarvam models only** — not available on the open-source models | | More than 128K of context in one request | [`glm5.2`](/api/getting-started/models/open-source/glm-5-2) | 512K context window — 4× the Sarvam chat models | | Image input with a conversational answer | [`gemma4`](/api/getting-started/models/open-source/gemma-4-31b) | The only chat model here that accepts images | | Document OCR or structured extraction | [Sarvam Vision](/api/getting-started/models/sarvam-vision) | Purpose-built for documents — better than `gemma4` for forms, invoices, PDFs | | Visible chain-of-thought (`reasoning_content`) | [`glm5.2`](/api/getting-started/models/open-source/glm-5-2) | Reasons before every answer and returns the trace — see [Reasoning output](#reasoning-output) | The two are not interchangeable. `glm5.2` reasons before every answer but cannot see images; `gemma4` sees images but never produces reasoning. Both support tool calling. ## Supported parameters These models accept the same OpenAI-compatible parameters as Sarvam's chat models, with the exceptions called out below. Support differs per model where the model itself lacks the capability. | Parameter | Type | Default | `glm5.2` | `gemma4` | Notes | | ------------------------ | ---------------- | ---------------------------------------------- | -------- | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | `messages` | array | — (required) | ✅ | ✅ | The conversation so far. At least one message. | | `model` | string | — (required) | ✅ | ✅ | `glm5.2` or `gemma4`. Must be an id returned by [`GET /v2/models`](/api-reference/open-source/models). | | `max_tokens` | integer | `2048` | ✅ | ✅ | Must be ≤ the model's context window. Estimated prompt tokens + `max_tokens` must fit inside the window, or the request is rejected with `422`. On `glm5.2` this budget also covers reasoning — see [Reasoning output](#reasoning-output). | | `temperature` | number | `0.5` if `reasoning_effort` is set, else `0.2` | ✅ | ✅ | `0`–`2`. Prefer altering this **or** `top_p`, not both. | | `top_p` | number | `1` | ✅ | ✅ | `>0`–`1` (`0` is rejected). Prefer altering this **or** `temperature`, not both. | | `stream` | boolean | `false` | ✅ | ✅ | Server-sent events. Errors after the stream starts arrive as an SSE `error` frame before `[DONE]`. | | `stop` | string or array | `null` | ✅ | ✅ | Up to 4 stop sequences; not included in the output. | | `n` | integer | `1` | ✅ | ✅ | `1`–`128`. You are billed for tokens across **all** choices — keep at `1` unless you need multiple completions. | | `seed` | integer | `null` | ⚠️ | ✅ | Beta, best-effort determinism. In current serving `gemma4` returns identical output for an identical seed and parameters; `glm5.2` does not. | | `frequency_penalty` | number | `0` | ✅ | ✅ | `-2`–`2`. Penalises tokens by existing frequency, reducing verbatim repetition. | | `presence_penalty` | number | `0` | ✅ | ✅ | `-2`–`2`. Penalises tokens already present, encouraging new topics. | | `reasoning_effort` | string | `null` | ✅ | ➖ | `low`, `medium`, or `high`. `glm5.2` reasons even when this is unset; `gemma4` accepts the parameter but never produces reasoning. See [Reasoning output](#reasoning-output). | | `response_format` | object | `null` | ✅ | ✅ | `{type: "text"}`, `{type: "json_object"}`, or `{type: "json_schema", json_schema: {...}}`. On `glm5.2`, turn thinking off — see [Structured Outputs](#structured-outputs). | | `tools` | array | `null` | ✅ | ✅ | OpenAI-compatible function calling. Max 128 tools; the model must support tools. | | `tool_choice` | string or object | `null` | ✅ | ✅ | `none`, `auto`, `required`, or a named function. Requires `tools` when set to `required` or a function name — otherwise `422`. | | `extra_body` | object | `{}` | ✅ | ✅ | Model-specific options forwarded unchanged. Must not repeat any field in this table or carry `stream_options`; a collision is rejected with `400`. Carries `chat_template_kwargs`, e.g. the thinking switch below. | | `image_url` content part | object | — | ❌ | ✅ | `glm5.2` returns `400`. `gemma4` accepts **base64 data URIs only** — remote URLs are rejected. See [Image input](#image-input). | Full request and response schema: [Open-Source Models API Reference](/api-reference/open-source/chat-completions). OpenAI-compatible fields not listed above — `logprobs`, `logit_bias`, `user`, and similar — are accepted but ignored. Unknown fields inside `extra_body` that the model backend does not recognise are rejected. **`wiki_grounding` is not part of the `/v2` schema.** For [Wikipedia grounding](/api/api-guides-tutorials/chat-completion/how-to/improve-response-factual-accuracy), use `sarvam-105b` on `/v1`. ### Reasoning output `glm5.2` is a reasoning model: it thinks before every answer, whether or not `reasoning_effort` is set. The chain-of-thought arrives in a separate `reasoning_content` field on the message — `content` holds only the final answer — and in streaming responses it arrives as `delta.reasoning_content` chunks before the answer starts. Reasoning tokens are billed as completion tokens and **count against `max_tokens`**, so a budget that is too small is consumed entirely by reasoning and the request returns `content: null` with `finish_reason: "length"`. To turn reasoning off — faster, cheaper, and the whole budget goes to the answer — send the thinking switch through `extra_body`: ```json { "extra_body": { "chat_template_kwargs": { "enable_thinking": false } } } ``` `gemma4` is not a reasoning model. It answers directly, never returns `reasoning_content`, and setting `reasoning_effort` on it has no effect. ### Structured Outputs `response_format` works on both models. Use `{"type": "json_schema", "json_schema": {...}}` to constrain output to a JSON Schema, or `{"type": "json_object"}` for the older JSON mode, which guarantees valid JSON but not a specific shape. **On `glm5.2`, turn thinking off whenever you set `response_format`** (as in the example below). With reasoning active, a JSON-constrained request returns `content: null` — the reasoning trace and the JSON grammar conflict, and no answer is produced. With `enable_thinking: false` the same request returns clean JSON reliably. #### Python ```python import json from sarvamai import SarvamAI client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY") response = client.open_source_models.chat_completions_v2( model="glm5.2", messages=[ {"role": "user", "content": "List two Indian cities with their populations."} ], max_tokens=500, extra_body={ "chat_template_kwargs": {"enable_thinking": False} }, response_format={ "type": "json_schema", "json_schema": { "name": "city_list", "strict": True, "schema": { "type": "object", "properties": { "cities": { "type": "array", "items": { "type": "object", "properties": { "name": {"type": "string"}, "population": {"type": "integer"}, }, "required": ["name", "population"], "additionalProperties": False, }, } }, "required": ["cities"], "additionalProperties": False, }, }, }, ) cities = json.loads(response.choices[0].message.content) print(cities) ``` #### JavaScript ```javascript import { SarvamAIClient } from "sarvamai"; const client = new SarvamAIClient({ apiSubscriptionKey: "YOUR_SARVAM_API_KEY", }); async function main() { const response = await client.openSourceModels.chatCompletionsV2({ model: "glm5.2", messages: [ { role: "user", content: "List two Indian cities with their populations.", }, ], extra_body: { chat_template_kwargs: { enable_thinking: false }, }, response_format: { type: "json_schema", json_schema: { name: "city_list", strict: true, schema: { type: "object", properties: { cities: { type: "array", items: { type: "object", properties: { name: { type: "string" }, population: { type: "integer" }, }, required: ["name", "population"], additionalProperties: false, }, }, }, required: ["cities"], additionalProperties: false, }, }, }, }); const cities = JSON.parse(response.choices[0].message.content); console.log(cities); } main(); ``` #### cURL ```bash curl -X POST https://api.sarvam.ai/v2/chat/completions \ -H "api-subscription-key: $SARVAM_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "glm5.2", "messages": [ {"role": "user", "content": "List two Indian cities with their populations."} ], "extra_body": { "chat_template_kwargs": { "enable_thinking": false } }, "response_format": { "type": "json_schema", "json_schema": { "name": "city_list", "strict": true, "schema": { "type": "object", "properties": { "cities": { "type": "array", "items": { "type": "object", "properties": { "name": {"type": "string"}, "population": {"type": "integer"} }, "required": ["name", "population"], "additionalProperties": false } } }, "required": ["cities"], "additionalProperties": false } } } }' ``` On the Python SDK, pass `extra_body` and `response_format` as direct keyword arguments to `chat_completions_v2` (as above). On `/v1/chat/completions`, the chat client does not expose `response_format` as a typed parameter — use `request_options={"additional_body_parameters": {"response_format": ...}}` instead; see [Structured Outputs](/api/api-guides-tutorials/chat-completion/overview#structured-outputs-with-json_schema). **Don't also ask for JSON in the prompt.** When `response_format` is set, an instruction like "Return JSON" in the message makes `gemma4` emit whitespace until it exhausts `max_tokens` — you are billed for the full budget and get no usable output. State the task in the prompt and let `response_format` handle the shape. ### Image input Only `gemma4` accepts images, and **only as base64 data URIs**. A remote `http://` or `https://` URL is rejected — the API does not fetch images on your behalf. Full example on the [Gemma 4 31B page](/api/getting-started/models/open-source/gemma-4-31b). ```json { "type": "image_url", "image_url": { "url": "data:image/png;base64," } } ``` Each image counts against the request-size cap once encoded. Prompt and request-size checks use estimates — text at roughly 4 bytes per token, plus a flat per-image allowance — rather than an exact tokenizer count. ### Tool schemas `function.parameters` must be a JSON Schema object; a property defined as a bare string is rejected. When `tool_choice` forces a call (`required` or a named function), schemas must be self-contained: non-local `$ref`, unresolved `$ref`, recursive schemas, or a combined expansion above 20,000 nodes across all tools in the request are rejected with `400`. ### Errors Every error uses the same envelope — `error.message`, `error.code`, and `error.request_id` (empty when the error is raised before a request id is assigned): ```json { "error": { "message": "Model 'gpt-4' not found.", "code": "not_found_error", "request_id": "20260729_0f1e2d3c-…" } } ``` | Status | Code | Raised when | | ------ | ---------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- | | `400` | `invalid_request_error` | Body fails validation, the model lacks a requested capability, the key lacks beta access to this endpoint, or `extra_body` collides with a validated field | | `402` | `insufficient_quota_error` | No credits available | | `403` | `invalid_api_key_error` | Missing or invalid API key | | `404` | `not_found_error` | Unknown model id, or a model not available to your key | | `413` | `invalid_request_error` | Request size exceeds the model's cap | | `422` | `unprocessable_entity_error` | Prompt + `max_tokens` exceeds the context window, `tool_choice` forced without `tools`, or the model rejected the request | | `429` | `rate_limit_exceeded_error` | Rate or concurrency limit; may carry `Retry-After` | | `500` | `internal_server_error` | Unhandled failure | | `502` | `model_call_failed` | The model call failed and could not be retried | | `503` | `model_overloaded` | Timeout, connection failure, or overload after automatic retries | Timeouts, connection failures, and overload are retried automatically before a `503` is returned. A request the model itself rejects is not retried and comes back as `422`. ## Support and expectations #### What Sarvam supports Sarvam supports the serving layer: availability, authentication, billing, rate limits, and API compatibility. If a request fails with a Sarvam error — an auth failure, a rate limit, a credits problem, a malformed response from our gateway — [contact us](/api/getting-started/help). #### What Sarvam does not support Sarvam does not tune, evaluate, or publish benchmarks for these models, and does not guarantee their output quality. A model's reasoning, factual accuracy, language coverage, and refusals are properties of the model itself rather than of Sarvam's serving. #### Availability and versioning These models may be updated, versioned, or withdrawn on a different schedule from Sarvam's own models. Do not assume one of these model IDs is a long-term stable contract — check the [changelog](/api/getting-started/changelog) before depending on one in production. #### Licensing and terms Each model carries its own open-source licence and acceptable-use terms, which apply to you as the end user in addition to Sarvam's terms. Review the model's licence before using it in a commercial product. ## Shared behaviour Everything below applies to both models: * **Endpoint** — `POST https://api.sarvam.ai/v2/chat/completions`. Full parameter reference: [Open-Source Models API Reference](/api-reference/open-source/chat-completions). * **Model discovery** — [`GET https://api.sarvam.ai/v2/models`](/api-reference/open-source/models) lists the model IDs the endpoint currently serves. * **Authentication** — `api-subscription-key: sk_xxx`. Invalid keys return `403 invalid_api_key_error`. See [Authentication](/api-reference/authentication). * **Beta access** — granted per key. Without access, the endpoint returns `400 invalid_request_error` before calling the model. See [Access to Beta APIs](/api-reference/beta-apis). * **Streaming** — set `"stream": true` for server-sent events. * **Billing** — metered per token against your Sarvam credits, reasoning tokens included. Per 1M tokens (input / cached input / output): `gemma4` ₹36.6 / ₹13.73 / ₹91.5, `glm5.2` ₹128.1 / ₹23.79 / ₹402.6. The two are not priced alike — `glm5.2` costs roughly 3.5× `gemma4`, and it reasons before every answer, so budget `max_tokens` deliberately. See [Pricing](/api/getting-started/pricing). * **Rate limits** — applied per API key. See [Credits & Rate Limits](/api/getting-started/ratelimits).