The models we build with and how we pick them.

For: hard decisions, document analysis, agent step planning

id: deepseek-v4

For: assistant on your knowledge, customer support, content translation

id: mistral-large-3

For: code generation and refactor, integrations, multilingual tasks

id: qwen3

For: chat, quality comparisons, tasks with light reasoning

id: glm

For: summaries, classification, extraction, cheap tasks at scale

id: gemma

For: writing code, automations, agent tooling, technical integrations

id: devstral

For: intent routing, lead classification, field extraction, quick answers

id: ministral

For: listing photo analysis, document reading, image description and tagging

id: qwen3-vl

For: photo analysis, image description and tagging, long-context tasks

id: kimi

For: mixed tasks, a backup chat engine, arena experiments

id: minimax

For: real-time interfaces, fast feedback, short answers

id: nemotron

For: complex reasoning, analytical tasks, hard decisions

id: gpt-oss

For: fallback reasoning, high-volume analysis, quality comparisons

id: cogito

For: long-document analysis, whole knowledge bases in one pass

id: gemini-3

For: semantic search, RAG, data privacy — the foundation of any AI knowledge base

id: bge-m3

For: on-prem deployments, full control over weights, no cloud dependency

id: llama-4

For: hardest reasoning, code work, quality-critical tasks

id: claude-opus

For: premium tasks where the client requires a specific provider, quality benchmarking

id: gpt-5

For: production assistants, code work, agents with a good quality-to-cost ratio

id: claude-sonnet

For: tasks needing current information, trend analysis, live research

id: grok-4

For: enterprise-grade RAG, company search, assistants with hard citation

id: command-a

For: on-prem deployments on modest hardware, edge tasks, low unit cost

id: phi-4