Overview
Kimchi is an OpenAI-compatible provider that serves open-weight models such as DeepSeek, GLM, MiniMax and Kimi. Bifrost delegates to the OpenAI implementation with standard parameter handling. Key characteristics:- OpenAI API compatibility - Same request and response format as OpenAI
- Full streaming support - Server-Sent Events with usage tracking
- Tool calling - Function definitions and tool choice in the OpenAI schema
- Reasoning - Reasoning models return
reasoning_content, which Bifrost maps to the message’sreasoning - Embeddings - OpenAI-compatible
/v1/embeddings - Responses API - Converted to Chat Completions
Supported Operations
Unsupported Operations (❌): Text Completions, Image Generation, Speech, Transcriptions, Files, and Batch return
UnsupportedOperationError. Kimchi does not serve the legacy /v1/completions endpoint.Setup & Configuration
Bifrost callshttps://llm.kimchi.dev/openai by default and sends the Kimchi API key as a bearer token. Bifrost appends the /v1/... paths itself. To create an API key, see the Kimchi documentation.
- Web UI
- config.json
- API
- Go SDK
- Navigate to Models > Model Providers. Look for Kimchi under Configured Providers. If it is missing, click on Add New Provider and select Kimchi.
- Click Add Key or edit an existing key.
- Set a name for your key.
- Paste your API key directly or use an environment variable (for example,
env.KIMCHI_API_KEY). - Set Allowed Models to All Models (default) or the specific model allowlist you want this key to serve.
- Save the provider configuration.
1. Chat Completions
Request Parameters
Kimchi supports the standard OpenAI chat completion parameters. For the full parameter reference and behavior, see OpenAI Chat Completions.Filtered Parameters
Removed before the request is sent, as for other OpenAI-compatible providers:predictionprompt_cache_key,prompt_cache_retention,prompt_cache_optionsverbositystoreweb_search_options
Reasoning
Reasoning models such asglm-5.3-flash return their thinking in reasoning_content, both in full responses and in stream deltas. Bifrost maps it to the message’s reasoning field and returns it under both reasoning and reasoning_content. Reasoning tokens are reported in usage.completion_tokens_details.reasoning_tokens.
2. Responses API
Converted internally to Chat Completions:previous_response_id and conversation are not carried over to the chat request, so send the full conversation history in input. Reasoning from chat responses becomes a reasoning output item.
3. List Models
Lists the models your key can use from Kimchi’s/v1/models, prefixed with kimchi/. When Kimchi returns a context_window for a model, Bifrost reports it as the model’s context length. Otherwise, context length and token limits come from Bifrost’s model catalog when it has an entry for the model.
4. Embeddings
Kimchi’s OpenAI-compatible/v1/embeddings endpoint, used through the standard embeddings interface. Embedding models such as bge-m3 are listed by /v1/models.
Streaming and Timeouts
How the default network configuration applies to Kimchi:- While a model produces no output (before the first token or mid-stream), Kimchi sends SSE keep-alive comments more often than Bifrost’s per-chunk idle timeout (
stream_idle_timeout_in_seconds, default 120). Every received byte resets that timer, so long silent stretches do not end the stream. - If a stream stays silent for a long time, Kimchi ends it with a final chunk that carries an
errorobject (type: "overloaded_error"). Bifrost returns it as a regular stream error, the same as a mid-stream error from any other provider. - Non-streaming requests are bounded by
default_request_timeout_in_seconds(default 300). Raise it in the provider’s network configuration if you run very long non-streaming generations. - When a model is busy, Kimchi can reject a request with a retryable 503
overloaded_error. Bifrost does not retry by default (max_retriesis 0), so setmax_retriesto 2 or more:
Unsupported Features
Caveats
Pricing
Pricing
Severity: Low
Behavior: Costs are computed from Bifrost’s model catalog
Impact: A Kimchi model without a catalog price is not costed
Fix: Add the model’s price as a pricing override
Parameter Filtering
Parameter Filtering
Severity: Low
Behavior: OpenAI-specific parameters filtered out
Impact: prompt_cache_key, verbosity, store and the other fields listed above are removed
Code: filterOpenAISpecificParameters

