Skip to content

Model IDs used in provider examples

Reviewed September 5, 2026. These are API identifiers, not display names. from_provider() adds the Instructor provider prefix shown below; direct SDK calls use the provider's model ID without that prefix.

Provider Instructor model string Provider reference
OpenAI openai/gpt-5.6-luna Model
Anthropic anthropic/claude-sonnet-5 Models
Google google/gemini-3.8-flash Model
Cohere cohere/command-a-03-2025 Models
Groq groq/llama-3.3-70b-versatile Model
Mistral mistral/mistral-small-latest Vision example
Fireworks fireworks/accounts/fireworks/models/kimi-k2p5 Model example
Cerebras cerebras/gpt-oss-120b Models
Writer writer/palmyra-x5 Models
xAI xai/grok-4.20-reasoning Model
Perplexity perplexity/sonar Model
DeepSeek deepseek/deepseek-v4-flash Migration
OpenRouter openrouter/google/gemini-3.8-flash Model
Together together/meta-llama/Llama-3.3-70B-Instruct-Turbo Models
Bedrock bedrock/anthropic.claude-sonnet-5 Model and regions
Vertex AI vertexai/gemini-3.8-flash Model guide

Request differences

  • Luna Chat Completions tool calls need reasoning_effort="none". Set output limits with max_completion_tokens. Use Responses for reasoning with tools.
  • Luna does not accept audio. Transcribe audio with gpt-4o-transcribe first.
  • Sonnet 5 adaptive thinking uses thinking={"type": "adaptive"} and tool_choice={"type": "auto"}; do not pass an extended-thinking token budget.
  • Fireworks requires the full accounts/fireworks/models/... name.
  • Bedrock and Vertex availability depends on region and account permissions. Check the provider reference before using a different region or inference profile.

Test coverage and older examples

tests/llm/test_current_doc_models.py makes real extraction requests. CI reports missing credentials as skips, not successful model validation. The Google test suite defaults to Gemini 3.8 Flash and accepts either a bare model ID or a google/-prefixed GOOGLE_GENAI_MODEL override.

Historical blog posts, recorded responses, and model-specific regression fixtures retain their original IDs. Haiku and GPT-4.1 tests remain as compatibility coverage. Databricks and Anyscale deployments, local Ollama models, and LiteLLM routes are deployment-specific; their names are not interchangeable with direct-provider IDs. This update does not certify those deployments or every provider mode.