Available models

We currently offer the following models, which may change or expand over time. Each is described along with model-specific parameters.

Model Name	Type	Modalities	Context (Tokens)	License
gpt-oss-120b	Chat + reasoning	Text, tool-calling	131,072	Apache 2.0
Qwen3.5-0.8B	Chat + reasoning	Text, tool-calling	262,144	Apache 2.0
Ministral-3-14B-Instruct-2512	Chat + vision	Text, image, tool-calling	262,144	Apache 2.0
Qwen3.5-122B-A10B-FP8	Chat + reasoning + vision	Text, image, tool-calling	245,760	Apache 2.0
Qwen3.6-35B-A3B-FP8	Chat + reasoning + vision	Text, image, tool-calling	262,144	Apache 2.0
GLM-OCR	Document OCR	PDF, DOCX, PPTX, XLSX, HTML, SVG, image to text	131,072	MIT
Qwen3-Embedding-8B	Embedding	Text to vector	32,768	Apache 2.0
Qwen3-VL-Reranker-2B	Reranking	Text, image to score	32,768	Apache 2.0
whisper-large-v3-turbo	Speech-to-Text	Audio to text	N/A (audio-based)	MIT

Tool-calling in this table reflects support on /v1/chat/completions. Support on /v1/responses is model-specific — see Responses API support for the exact per-model breakdown.

Picking models

• For complex text-centric workloads and advanced automations when precision and vast knowledge are required
  use gpt-oss-120b.
• For high-throughput, cost-sensitive tasks that don't require vision (e.g. simple Q&A, routing, classification, and batch processing)
  use Qwen3.5-0.8B.
• For broad, scalable, cost-conscious chat and basic multimodal (text + image) workflows
  use Ministral-3-14B-Instruct-2512.
• For large-scale reasoning and vision tasks where high model capacity is required
  use Qwen3.5-122B-A10B-FP8.
• For workloads that require long context windows with reasoning and vision support at lower cost
  use Qwen3.6-35B-A3B-FP8.
• Special-purpose applications
  • For extracting text from PDF, DOCX, PPTX, XLSX, HTML, and image documents — including scanned invoices, contracts, and forms –
    use GLM-OCR
  • For all use cases involving search, recommendation, clustering, or knowledge graph building
    use Qwen3-Embedding-8B
  • For any audio transcription or voice-command needs
    use whisper-large-v3-turbo
  • To improve RAG retrieval precision by adding it as a second-pass reranker after vector search
    use Qwen3-VL-Reranker-2B

For more details and additional tips have a look at the usage examples and guides.

tip

For optimal ROI, start with the smallest model that meets your quality bar—then upgrade only where you need additional depth or reliability! "Small" here refers to the number of parameters a model has, which often correlates with cost and compute requirements. The "context size" of a model (measured in tokens) is also a crucial factor to consider, as it determines how much information the model can process at once, which can also impact performance and suitability for certain tasks. So when it comes to model selection, it's not just about picking the most powerful model, but rather the one that best fits your specific use case.

Please have a look at the following pages to gather more information about a specific model:

📄️Ministral-3-14B-Instruct-2512

Detailled information on Ministral-3-14B-Instruct-2512

📄️Qwen3-Embedding-8B

Detailled information on Qwen3-Embedding-8B

📄️gpt-oss-120b

Detailled information on gpt-oss-120b

📄️Whisper-Large-V3-Turbo

Detailed information about Whisper-Large-V3-Turbo

📄️Qwen3.5-122B-A10B-FP8

Detailled information on Qwen3.5-122B-A10B-FP8

📄️Qwen3.6-35B-A3B-FP8

Detailled information on Qwen3.6-35B-A3B-FP8

📄️GLM-OCR

Detailed information about GLM-OCR

📄️Qwen3.5-0.8B

Detailed information about Qwen3.5-0.8B

📄️Qwen3-VL-Reranker-2B

Detailed information about Qwen3-VL-Reranker-2B