Migrating Chat Completion from Together AI to Sarvam

Last updated: October 23, 2026

This guide covers moving Together AI Chat Completions calls for open-weight models to Sarvam’s /v2 API.

In short: change base_url, api_key, and model. Replace reasoning: {"enabled": false} with reasoning_effort: "none", send logprobs as true instead of a number, and remove Together-only fields such as safety_model: Sarvam returns 400 for request fields it doesn’t support. Remember that 403 now means an auth error, not a context-length error.

Migration checklist

Create an API key in the Sarvam dashboard.
Change base_url to https://api.sarvam.ai/v2 and api_key to your Sarvam key.
Replace each model ID using the model ID table.
Replace reasoning: {"enabled": false} with reasoning_effort: "none" (Reasoning).
Remove Together-only fields: safety_model, compliance, echo, and others (list).
Send logprobs as true, with the count in top_logprobs.
Update error handling: 403 now means an invalid key, not a prompt that’s too long.
Set max_tokens explicitly, then run your tests and watch for 400 responses that name a field.

Quick start

	Together AI	Sarvam
Base URL	`https://api.together.ai/v1`	`https://api.sarvam.ai/v2`
Auth header	`Authorization: Bearer <key>`	`api-subscription-key: <key>` or `Authorization: Bearer <key>`
Model ID	`moonshotai/Kimi-K3`	`kimi-k3` (see the model ID table)

Before (Together AI)

After (Sarvam)

from together import Together
client = Together(api_key="YOUR_TOGETHER_API_KEY")
response = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-V4-Pro-0813",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Suggest a name for a coffee shop in Bengaluru."},
    ],
    max_tokens=2000,
    reasoning={"enabled": False},
)
print(response.choices[0].message.content)

Reasoning

Together AI	Sarvam	Note
`reasoning_effort: "low"` / `"high"` / `"max"`	Same	Each model’s default is listed under Reasoning.
`reasoning_effort: "medium"`	Use `"low"`, `"high"`, or `"max"`	Accepted on some models and returns `400` on others, such as `glm5.3`.
`reasoning: {"enabled": false}`	`reasoning_effort: "none"`	The `reasoning` object returns `400`. `"none"` turns thinking off on `deepseekv4-pro`, `mimo-v2.6-flash`, and `mimo-v2.6-pro`. `kimi-k3` and `glm5.3-flash` always think.
`reasoning: {"enabled": true}`	Remove	Reasoning models think by default.
`chat_template_kwargs`	Same	Passed to the model. `{"enable_thinking": false}` turns thinking off on the MiMo models.
`choices[].message.reasoning` (DeepSeek V4 Pro)	`choices[].message.reasoning_content`	Sarvam uses `reasoning_content` for every model, and `delta.reasoning_content` when streaming.
`choices[].message.reasoning_content` (Kimi K3, GLM)	Same

Together-only fields to remove

Sarvam validates request fields and returns 400 invalid_request_error for fields it doesn’t support, so leftover Together options fail loudly instead of being ignored.

Together field	What to do on Sarvam
`reasoning` object	Replace with `reasoning_effort` (see Reasoning).
`safety_model`	Remove. Run moderation as a separate step before or after the call.
`compliance`	Remove.
`context_length_exceeded_behavior`	Remove. Sarvam returns an error when the prompt plus `max_tokens` is longer than the context window, as Together does by default.
`echo`	Remove. Keep the prompt on your side.
`response_format: {"type": "regex"}`	Use `response_format` with a `json_schema`, and put the pattern in a string property’s `pattern`.
`logprobs: 5` (a number)	Use `logprobs: true` with `top_logprobs: 5`.
`functions`, `function_call`	Use `tools` and `tool_choice`. The legacy fields return `400`.

Parameter concept map

Together AI	Sarvam	Note
`messages`, `max_tokens`, `temperature`, `top_p`, `top_k`, `stop`, `seed`	Same	Set `max_tokens` explicitly: when it’s omitted, generation isn’t capped at a small default.
`repetition_penalty`, `frequency_penalty`, `presence_penalty`	Same
`min_p`	Not supported	Remove it. The open-weight models accept only `0`, which is the same as leaving it out.
`tools`, `tool_choice`, `parallel_tool_calls`	Same	`parallel_tool_calls: false` is enforced on every model; see V2 feature support.
`response_format` (`json_schema`, `json_object`)	Same
`n`	Same	Supported on most models. `temperature` must be above `0` when `n` > 1.
`stream: true`	Same	As on Together, the last chunk before `[DONE]` carries `usage` with an empty `choices` array; keep skipping it.
`image_url` content parts	Same	Sarvam accepts inline base64 data URIs only, not remote image URLs, on models with image input. See each model page.
`video_url` content parts	Same, as a base64 data URI	On the MiMo models and `kimi-k3` (beta, short clips). Sarvam doesn’t fetch video URLs.
`input_audio` content parts	Same, on the MiMo models	`wav` or `mp3`. For FLAC, OGG, or Opus, send an `audio_url` part with a data URI. See Media input.
`finish_reason: "eos"`	`finish_reason: "stop"`

Responses API

Together doesn’t offer a Responses API. Sarvam does: if you want the OpenAI Responses format, see the Responses API guide. It’s stateless, like Chat Completions.

What’s different by design

No dedicated endpoints to manage. Models that are dedicated-only on Together, such as Gemma 4 31B, are called by model ID on Sarvam.
Strict request validation. Unsupported fields return 400 with the field named, instead of being dropped silently. gemma4 is less strict: it accepts some fields, such as prediction and logit_bias, and ignores them.
One reasoning field. Reasoning comes back in reasoning_content for every model.
Different meaning for 403. On Sarvam it’s an auth error. See the status code table.

Platform features without a direct equivalent

Together feature	On Sarvam
Batch API	Not available. Send requests concurrently within your rate limits.
Fine-tuning and LoRA adapters	Not available on the `/v2` API.
Dedicated endpoints and provisioned throughput	Not needed for these models. Contact Sarvam for capacity needs.
Llama Guard moderation (`safety_model`)	Not available. Run moderation separately.

Full example

A wrapper that turns thinking off, asks for log probabilities, and maps status codes:

from together import Together
client = Together(api_key="YOUR_TOGETHER_API_KEY")
def ask(prompt: str) -> str:
    response = client.chat.completions.create(
        model="deepseek-ai/DeepSeek-V4-Pro-0813",
        messages=[{"role": "user", "content": prompt}],
        max_tokens=2000,
        reasoning={"enabled": False},
        logprobs=3,
    )
    return response.choices[0].message.content

Common mistakes

Treating 403 as a context-length error

On Sarvam, 403 means the API key is missing or invalid. Don’t trim the prompt and retry on 403; check the key instead.

Keeping the organisation prefix in the model ID

moonshotai/Kimi-K3 returns 404 on Sarvam. Use kimi-k3. Model IDs are case-sensitive. See the model ID table.

Sending reasoning: {enabled: false}

Sarvam returns 400. Use reasoning_effort: "none" on models that can turn thinking off.

Sending logprobs as a number

Sarvam expects logprobs: true and the count in top_logprobs.

Sending images or video by URL

Sarvam accepts images, audio, and video only as inline base64 data URIs; URLs return 400. Download the file and encode it before sending. See Media input.