> 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. # Migrating Chat Completion from Fireworks AI to Sarvam > Request fields, reasoning, Responses API, and Fireworks-only features to change when moving chat traffic from Fireworks AI to Sarvam's /v2 API, with before/after code. *Last updated: October 23, 2026* This guide covers moving Fireworks AI Chat Completions and Responses calls for open-weight models to Sarvam's [`/v2` API](/api/token-factory/chat-completions). > **Tip** > > **In short:** change `base_url`, `api_key`, and `model`. Reasoning comes back in > `reasoning_content` on both platforms. Remove Fireworks-only fields such as > `prompt_truncate_len` and `perf_metrics_in_response`, because Sarvam returns `400` for > request fields it doesn't support. Set `max_tokens` explicitly, and send the full history > on every Responses call. ## Migration checklist 1. Create an API key in the [Sarvam dashboard](https://indus.sarvam.ai/key-management?ref=docs). 2. Change `base_url` to `https://api.sarvam.ai/v2` and `api_key` to your Sarvam key. 3. Replace each model ID using the [model ID table](/api/token-factory/migrations/from-fireworks/overview#model-ids). Fast, US, and deployment IDs map to the same Sarvam ID. 4. Use string values for `reasoning_effort` (`low`, `high`, `max`, or `none`) ([Reasoning](#reasoning)). 5. Remove Fireworks-only fields: `prompt_truncate_len`, `context_length_exceeded_behavior`, `perf_metrics_in_response`, and others ([list](#fireworks-only-fields-to-remove)). 6. Set `max_tokens` explicitly: only `glm5.3` defaults to `2048` on Sarvam. 7. Trim long conversations yourself: Sarvam doesn't truncate prompts. 8. If you use the Responses API, follow [Migrating the Responses API](/api/token-factory/migrations/from-fireworks/responses-api). 9. Run your tests and watch for `400` responses that name a field. ## Quick start | | Fireworks AI | Sarvam | | ----------- | --------------------------------------- | ----------------------------------------------------------------------------------------------------- | | Base URL | `https://api.fireworks.ai/inference/v1` | `https://api.sarvam.ai/v2` | | Auth header | `Authorization: Bearer ` | `api-subscription-key: ` or `Authorization: Bearer ` | | Model ID | `accounts/fireworks/models/kimi-k3` | `kimi-k3` (see the [model ID table](/api/token-factory/migrations/from-fireworks/overview#model-ids)) | #### Before (Fireworks) #### Python (OpenAI SDK) ```python from openai import OpenAI client = OpenAI( api_key="YOUR_FIREWORKS_API_KEY", base_url="https://api.fireworks.ai/inference/v1", ) response = client.chat.completions.create( model="accounts/fireworks/models/kimi-k3", messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Suggest a name for a coffee shop in Bengaluru."}, ], reasoning_effort="low", extra_body={"prompt_truncate_len": 100000}, ) print(response.choices[0].message.content) ``` #### JavaScript (OpenAI SDK) ```javascript import OpenAI from "openai"; const client = new OpenAI({ apiKey: "YOUR_FIREWORKS_API_KEY", baseURL: "https://api.fireworks.ai/inference/v1", }); const response = await client.chat.completions.create({ model: "accounts/fireworks/models/kimi-k3", messages: [ { role: "system", content: "You are a helpful assistant." }, { role: "user", content: "Suggest a name for a coffee shop in Bengaluru." }, ], reasoning_effort: "low", prompt_truncate_len: 100000, }); console.log(response.choices[0].message.content); ``` #### cURL ```bash curl https://api.fireworks.ai/inference/v1/chat/completions \ -H "Authorization: Bearer YOUR_FIREWORKS_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "accounts/fireworks/models/kimi-k3", "messages": [{"role": "user", "content": "Suggest a name for a coffee shop in Bengaluru."}], "reasoning_effort": "low", "prompt_truncate_len": 100000 }' ``` #### After (Sarvam) #### Python (OpenAI SDK) ```python from openai import OpenAI client = OpenAI( api_key="YOUR_SARVAM_API_KEY", base_url="https://api.sarvam.ai/v2", ) response = client.chat.completions.create( model="kimi-k3", messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Suggest a name for a coffee shop in Bengaluru."}, ], reasoning_effort="low", max_tokens=2000, ) print(response.choices[0].message.content) ``` #### JavaScript (OpenAI SDK) ```javascript import OpenAI from "openai"; const client = new OpenAI({ apiKey: "YOUR_SARVAM_API_KEY", baseURL: "https://api.sarvam.ai/v2", }); const response = await client.chat.completions.create({ model: "kimi-k3", messages: [ { role: "system", content: "You are a helpful assistant." }, { role: "user", content: "Suggest a name for a coffee shop in Bengaluru." }, ], reasoning_effort: "low", max_tokens: 2000, }); console.log(response.choices[0].message.content); ``` #### Python (Sarvam SDK) ```python from sarvamai import SarvamAI client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY") response = client.chat.completions_v2( model="kimi-k3", messages=[ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Suggest a name for a coffee shop in Bengaluru."}, ], reasoning_effort="low", max_tokens=2000, ) print(response.choices[0].message.content) ``` #### JavaScript (Sarvam SDK) ```javascript import { SarvamAIClient } from "sarvamai"; const client = new SarvamAIClient({ apiSubscriptionKey: "YOUR_SARVAM_API_KEY" }); const response = await client.chat.completionsV2({ model: "kimi-k3", 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_effort: "low", }); console.log(response.choices[0].message.content); ``` #### cURL ```bash curl https://api.sarvam.ai/v2/chat/completions \ -H "api-subscription-key: YOUR_SARVAM_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "kimi-k3", "messages": [{"role": "user", "content": "Suggest a name for a coffee shop in Bengaluru."}], "reasoning_effort": "low", "max_tokens": 2000 }' ``` ## Reasoning | Fireworks | Sarvam | Note | | ----------------------------------------------------- | --------------------------------- | -------------------------------------------------------------------------------------------------------------------------- | | `reasoning_effort: "low"` / `"high"` / `"max"` | Same | Each model's default is listed under [Reasoning](/api/token-factory/models#reasoning). | | `reasoning_effort: "medium"` / `"xhigh"` | Use `"low"`, `"high"`, or `"max"` | Accepted on some models and returns `400` on others, such as `glm5.3`. | | `reasoning_effort: "none"` or `false` | `reasoning_effort: "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_effort` as a token budget (`1024` or more) | Use an effort level | Integer budgets above `100` return `400`. | | `thinking: {"type": "disabled"}` | Same, via `extra_body` | Works on the models that can turn thinking off. To control depth, use `reasoning_effort`. | | `reasoning_history` | Remove | To keep reasoning across turns, send the earlier assistant `reasoning_content` back in `messages`. | | `choices[].message.reasoning_content` | Same | Also `delta.reasoning_content` when streaming. | > **Note** > > On Fireworks, a `json_schema` response format turns reasoning off. On Sarvam, reasoning > models keep reasoning when you use `response_format`, so leave room in `max_tokens`, or > turn thinking off on models that support it. ## Fireworks-only fields to remove Sarvam validates request fields and returns `400 invalid_request_error` with `Unsupported parameter(s)` for fields it doesn't support, so leftover Fireworks options fail loudly instead of being ignored. | Fireworks field | What to do on Sarvam | | --------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------- | | `prompt_truncate_len` | Remove. Trim old messages yourself before sending. | | `context_length_exceeded_behavior` | Remove. Sarvam returns an error when the prompt plus `max_tokens` is longer than the context window, instead of truncating. | | `perf_metrics_in_response` | Remove. Measure latency on your side. Sarvam returns token counts in `usage`. | | `response_format: {"type": "grammar"}` | Use `response_format` with a `json_schema`. Sarvam supports `text`, `json_object`, and `json_schema`. | | `typical_p`, `mirostat_target`, `mirostat_lr` | Remove. Use `temperature`, `top_p`, or `top_k`. | | `raw_output`, `echo`, `return_token_ids`, `prompt_token_ids`, `safe_tokenization` | Remove. Not available. | | `prediction`, `logit_bias` | Remove. Not supported. | | `reasoning_history` | Remove (see [Reasoning](#reasoning)). | | Model IDs with `accounts/.../deployments/...` | Use the Sarvam model ID. There are no deployments to manage. | ## Parameter concept map | Fireworks | Sarvam | Note | | ------------------------------------------------------------- | -------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `messages`, `temperature`, `top_p`, `top_k`, `stop`, `seed` | Same | | | `max_tokens` (defaults to `2048`) | `max_tokens` | Set it explicitly. When it's omitted on Sarvam, generation isn't capped at `2048` (except on `glm5.3`). | | `frequency_penalty`, `presence_penalty`, `repetition_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 | `tool_choice: "any"` is Fireworks-only; use `"required"`. | | `response_format` (`json_object`, `json_schema`) | Same | | | `logprobs`, `top_logprobs` | Same | `logprobs` must be `true` or `false`. Not supported on `glm5.3`. | | `n` | Same | Supported on most models. `temperature` must be above `0` when `n` > 1. | | `prompt_cache_key`, `user` | Accepted, except `prompt_cache_key` on `gemma4`, which returns `400` | For multi-turn caching on `gemma4`, send the [`x-session-affinity`](/api/token-factory/chat-completions#session-affinity-for-multi-turn-conversations) header. | | `service_tier: "priority"`, Fast routers | Not needed | Sarvam has a single tier per model. | | `stream: true` | Same | Usage arrives in the final chunk. | | `image_url` content parts | Same | Inline base64 data URIs only, on models with image input. See each [model page](/api/token-factory/models). | | `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](/api/token-factory/migrations/from-fireworks/overview#media-input-and-request-size). | | `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. | ## Responses API Fireworks stores responses and chains them with `previous_response_id`; Sarvam's Responses API is stateless, so you send the history in `input` instead. For base URL and field changes, reasoning items, function tools, and server-side tools, see [Migrating the Responses API from Fireworks](/api/token-factory/migrations/from-fireworks/responses-api). ## What's different by design * **No deployments to manage.** Models that need an on-demand deployment on Fireworks, such as Gemma 4 31B and DeepSeek V4 Pro, 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. * **No silent truncation.** Sarvam rejects a prompt that doesn't fit, rather than cutting old messages or lowering `max_tokens` for you. * **Stateless Responses.** Your app keeps the conversation history, as with Chat Completions. ## Platform features without a direct equivalent | Fireworks feature | On Sarvam | | -------------------------- | -------------------------------------------------------------------------------------------------------------- | | Batch inference jobs | Not available. Send requests concurrently within your [rate limits](/api/getting-started/pricing#rate-limits). | | Fine-tuned and LoRA models | Not available on the `/v2` API. | | On-demand deployments | Not needed. Every model is served by ID. | | US-only endpoint | Not available. | ## Full example A wrapper with retries on `429` and `503`, which Fireworks returns when it sheds load and Sarvam returns when a model is overloaded: #### Before (Fireworks) ```python from openai import OpenAI client = OpenAI( api_key="YOUR_FIREWORKS_API_KEY", base_url="https://api.fireworks.ai/inference/v1", ) def ask(prompt: str) -> str: response = client.chat.completions.create( model="accounts/fireworks/models/deepseek-v4p1-flash", messages=[{"role": "user", "content": prompt}], reasoning_effort="low", extra_body={"context_length_exceeded_behavior": "truncate"}, ) return response.choices[0].message.content ``` #### After (OpenAI SDK) ```python import time from openai import OpenAI, APIStatusError client = OpenAI( api_key="YOUR_SARVAM_API_KEY", base_url="https://api.sarvam.ai/v2", ) def ask(prompt: str) -> str: for attempt in range(3): try: response = client.chat.completions.create( model="deepseekv4.1-flash-1017", messages=[{"role": "user", "content": prompt}], reasoning_effort="low", max_tokens=2000, ) return response.choices[0].message.content except APIStatusError as e: if e.status_code in (429, 503): time.sleep(2 ** attempt) # back off, then retry continue raise # 400, 402, 403 and 404 won't succeed on retry raise RuntimeError("Model is busy. Try again later.") ``` #### After (Sarvam SDK) ```python import time from sarvamai import SarvamAI from sarvamai.core.api_error import ApiError client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY") def ask(prompt: str) -> str: for attempt in range(3): try: response = client.chat.completions_v2( model="deepseekv4.1-flash-1017", messages=[{"role": "user", "content": prompt}], reasoning_effort="low", max_tokens=2000, ) return response.choices[0].message.content except ApiError as e: if e.status_code in (429, 503): time.sleep(2 ** attempt) # back off, then retry continue raise # 400, 402, 403 and 404 won't succeed on retry raise RuntimeError("Model is busy. Try again later.") ``` ## Common mistakes #### Keeping the accounts/fireworks/models/ prefix `accounts/fireworks/models/kimi-k3` returns `404` on Sarvam. Use `kimi-k3`. See the [model ID table](/api/token-factory/migrations/from-fireworks/overview#model-ids). #### Relying on the 2048-token max_tokens default Fireworks caps output at `2048` tokens when `max_tokens` is omitted. On Sarvam only `glm5.3` does, so long answers can run much longer. Set `max_tokens` explicitly, leaving room for reasoning tokens. #### Expecting long prompts to be truncated Sarvam returns an error instead of truncating. Trim the conversation before sending. #### Chaining Responses calls with previous_response_id Sarvam doesn't store responses. Keep the history in your app and send it in `input`, as in the [Responses example](#responses-api). #### Leaving Fireworks fields in extra_body Fields like `prompt_truncate_len` and `perf_metrics_in_response` are often passed through `extra_body`. Sarvam returns `400` with the field name; remove each one it names. ## See also * [Migration overview](/api/token-factory/migrations/from-fireworks/overview): base URL, auth, and model IDs * [Chat Completion overview](/api/token-factory/chat-completions) * [Responses API guide](/api/token-factory/responses-api) * [Open-Weight Models](/api/token-factory/models) * [Errors & Troubleshooting](/api/getting-started/errors-troubleshooting)