Supported Models

The table below lists the models currently available in the Xinference catalog. The catalog is updated regularly as new models are evaluated and approved.

Tip

The live catalog is always available via the API: GET /v1/deployable-models

Large Language Models (LLM)

Model Name Display Name Provider Size Quantization Tags
qwen2.5-instruct Qwen2.5 Instruct Alibaba 7B, 14B, 32B, 72B int4, int8, none chat, multilingual, code
qwen2.5-coder-instruct Qwen2.5 Coder Alibaba 7B, 14B int4, int8 chat, code
llama-3.1-instruct Llama 3.1 Instruct Meta 8B, 70B int4, int8, none chat
llama-3.2-instruct Llama 3.2 Instruct Meta 1B, 3B int4, none chat, edge
mistral-instruct Mistral Instruct Mistral AI 7B int4, int8 chat
mixtral-instruct Mixtral Instruct Mistral AI 8×7B int4 chat
deepseek-r1-distill DeepSeek R1 Distill DeepSeek 7B, 14B int4 chat, reasoning
gemma-2-instruct Gemma 2 Instruct Google 9B, 27B int4, int8 chat

Embedding Models

Model Name Display Name Provider Dimension Tags
bge-m3 BGE M3 BAAI 1024 embedding, multilingual
bge-large-en-v1.5 BGE Large EN v1.5 BAAI 1024 embedding, english
text-embedding-3-small Text Embedding 3 Small OpenAI-compat 1536 embedding
gte-qwen2-7b-instruct GTE Qwen2 7B Alibaba 3584 embedding, multilingual

Instance Type Mapping

Xinference automatically selects the most cost-effective GPU instance for each model. The general mapping:

Model Size Typical Instance
≤ 7B (int4) Single A10G (24 GB)
7B–14B (int4) Single A10G or A100 40 GB
14B–32B (int4) A100 40 GB
70B+ (int4) 2× A100 80 GB
Embedding CPU or T4

Actual selection depends on availability and model format. See Deployment Overview → for details.