Skip to main content
Models on SGL Grid are served by independent TEE nodes. Availability is live — a model is callable whenever at least one attested node is serving it.

​
Browsing models

The dashboard’s Models view lists every model currently served, with its provider, size, and per-request price. You can search and sort (e.g. cheapest first). Programmatically, list models via the API:
Each entry includes the model id (use it as model in your request), how many nodes serve it, and its price.

​
Model families

The grid serves several kinds of model, shown as coloured categories in the dashboard:
  • Language — general chat / instruct models (Llama, Qwen, Gemma, Mistral, Phi …).
  • Code — programming-tuned models (Qwen Coder) with OpenAI tool-calling.
  • Reasoning — chain-of-thought models (DeepSeek R1) that “think” before answering; their reasoning shows in a collapsible box in chat.
  • Embeddings — text-to-vector models for search and RAG, plus release-gated EmbeddingGemma 2 vectors for ordered text, image, audio, video, and mixed inputs. These use POST /v1/embeddings and are not chat models — see Multimodal embeddings.
  • System One — typed decision models such as Laya (convaiinnovations/laya). These use POST /v1/systemone, not chat completions — see Laya / System One.
  • Speech to text — release-gated audio transcription. whisper-1 appears only after the public gates and an exact ready provider pass. It uses metadata reservation plus client-sealed JSON, not Chat Completions or multipart upload — see Speech-to-text transcription.
Language, code and reasoning models are called through the same chat completions endpoint; embeddings, System One, and transcription have their own endpoints.

​
Choosing a model

  • Smaller models (e.g. 3B-class) are cheapest and fastest — great for classification, extraction, drafting, and agent steps.
  • Larger models trade cost/latency for capability.
  • Reasoning models are best for math, logic and multi-step planning.
  • System One models are best for typed choices, scoring, and decision steps where you want a structured answer.
  • Availability matters — a model with more serving nodes is more resilient. Prefer models with redundancy for production.

​
OpenAI compatibility

Chat and embeddings use their documented OpenAI-compatible shapes. Transcription deliberately does not accept OpenAI multipart input; use the official SGL SDK helper. See the API guide.
The catalog grows as operators bring more models online. Want a model served? Run a node and serve it, or ask the community.