API & Models
Available Models
Dedicated Gateway serves as a proxy to the full IG1 AI catalog.
LLM Models
| Model ID | Model Name | Capabilities | Max Context |
|---|---|---|---|
kimi-k26-instant | Kimi K2.6 1T A32B Instant | Standard, High-velocity coding, code review, architectural decisions | 256K tokens |
kimi-k26-thinking | Kimi K2.6 1T A32B Thinking | Deep debugging, algorithm design, complex system modeling | 256K tokens |
kimi-k26-thinking-preserve | Kimi K2.6 1T A32B Thinking Preserve | Autonomous agents with preserved reasoning traces across turns | 256K tokens |
qwen35-122b-a10b-instruct-general | Qwen3.5-122B-A10B | Fast, accurate, and ready for any mixed workload. | 256K tokens |
qwen35-122b-a10b-instruct-creative | Qwen3.5-122B-A10B Creative | Perfect for copywriting, storytelling, and brainstorming sessions that need a spark. | 256K tokens |
qwen35-122b-a10b-thinking-general | Qwen3.5-122B-A10B Thinking | Engages deep reasoning | 256K tokens |
qwen35-122b-a10b-thinking-coding | Qwen3.5-122B-A10B Thinking Coder | pair programmer for algorithms, refactoring, and system design | 256K tokens |
LLM models support images up to 4096 × 4096 pixels. However, processing images high-resolution images (more than 4K) can significantly increase:
- Computing time: Varies based on resolution and image complexity
- Token consumption: Higher resolutions use more input tokens
Recommendation: For optimal performance and cost efficiency, resize images to 2K maximum (2560 × 1440 pixels) before sending.
Select the right LLM model
Kimi K2.6
Choose it when you're building software, running multi-agent workflows, or tackling complex engineering tasks that demand frontier-level reasoning.
Kimi K2.6 delivers Claude Opus 4.6-class performance with a focus on agentic coding. It is perfect to architects solutions, orchestrates multi-agent collaborations, and thinks several steps ahead.
All Kimi K2.6 virtual models carry a suffix (-instant, -thinking, -thinking-preserve) to indicate that the request is altered by the IG1 AI routing layer. None of them provide direct, unmodified access to the raw model.
Kimi K2.6 is a thinking model by default. The -instant suffix alters the request to force the model to not produce thinking traces and jump to the final response directly, like a non thinking model.
Kimi K2.6 supports two distinct thinking behaviors:
Interleaved thinking (default in kimi-k26-thinking)
- The model can produce reasoning, then call a tool, then continue reasoning — all within a single conversational turn.
- If the client sends back the reasoning content during this interleaved flow, the model can continue its thinking trace using the tool results. This avoids forcing the model to restart its thinking trace from scratch, using a new tool response as a fresh starting point. This greatly enhances model coherence and reasoning quality; therefore, clients are strongly encouraged to echo back thinking traces. Thinking traces from previous conversational turns—where the assistant has already produced its final answer—are automatically dropped and excluded from input token counts.
- This is already the default behavior in the chat template and works as long as the client returns the reasoning content.
Preserved thinking (kimi-k26-thinking-preserve)
- The model preserves its full reasoning trace across multiple conversational turns.
- This prevents an autonomous agent from emitting a hypothesis, invalidating it, and then re-emitting the same hypothesis in the next turn because it no longer "remembers" it.
- Over multiple turns, this reduces redundant thinking and makes the agent more performant.
- Purpose-built for autonomous agents and long-running workflows.
Qwen 3.5 122B
Choose it when you need a reliable, high-performance model that handles everything from creative writing to visual analysis — without breaking the budget.
Qwen 3.5 122B is the Swiss Army knife of our model lineup. It does it all, and it does it cost-effectively. With four tuned variants, you get precisely the right personality for every task. :::
The Creative and Thinking Coder variants enforce specific sampling parameters on the backend. If you need full control over sampling, use the base instruct-general or thinking-general models instead.
Details and usage example available under API documentation.
Use your dedicated URL Endpoint instead of api.ig1.ai
Embedding & Reranking models
| Model ID | Model Name | Dimensions | Max Context | IG1 AI Model Name |
|---|---|---|---|---|
qwen3-vl-embedding-8b | Qwen3 VL Embedding | 128 – 4096 (configurable) | 32K tokens | Qwen3-VL-Embedding 8B |
bge-m3 | BGE-M3 | 1024 (fixed) | 8K tokens | BGE-m3 |
bge-reranker-v2-m3 | BGE-Reranker-v2-M3 | N/A (reranker) | BGE-Reranker-v2-m3 |
Image models
| Model ID | Model Name | Capabilities | IG1 AI Model Name |
|---|---|---|---|
qwen-image | Qwen Image | Image generation | Qwen Image |
qwen-image-pe | Qwen Image PE | Image generation with prompt enhancement | Qwen Image PE |
qwen-image-edit | Qwen Image Edit | Image editing | Qwen Image Edit |
qwen-image-edit-pe | Qwen Image Edit PE | Image editing with prompt enhancement | Qwen Image Edit PE |
Virtual Models
Virtual Models are custom named models on top of IG1 AI models. By default IG1 teams deploy following models on each dedicated gateways.
| Virtual Model ID | IG1 AI Model ID |
|---|---|
coding-max | kimi-k26-thinking |
coding-max-thinking | kimi-k26-thinking-preserve |
general-max | kimi-k26-instant |
general-max-thinking | kimi-k26-thinking |
coding-pro-thinking | qwen35-122b-a10b-thinking-coding |
general-pro | qwen35-122b-a10b-instruct-general |
general-pro-thinking | qwen35-122b-a10b-thinking-general |
image-pro | qwen-image-pe |
image-pro-edit | qwen-image-edit-pe |
API Supported endpoints
Dedicated Gateway supports following API endpoints from IG1 AI API documentation:
- GET
/v1/models - POST
/v1/chat/completions - POST
/v1/embeddings - POST
/cohere/v2/embed - POST
/cohere/v2/rerank - POST
/v1/images/generations - POST
/v1/images/edits - POST
/tokenize - GET
/pricing
Please refer to api reference for all details.
Replace API reference endpoint URL from example by your own Dedicated Gateway endpoint.
Example:
Replace https://api.ig1.ai by https://api.demo.ig1.ai if you use the Demo Endpoint.
API Unsupported endpoints
Dedicated Gateway do not support following endpoints:
- POST
/document - GET
/usage - GET
/usage/history
You should use your API Master Key and direct request to https://api.ig1.ai instead if you need to call these endpoints.