Skip to content

Latest commit

 

History

History
87 lines (63 loc) · 2.97 KB

providers.md

File metadata and controls

87 lines (63 loc) · 2.97 KB

Model Providers

Overview

The OpenCode Agent Loop supports multiple model providers for agent orchestration. Models are organized into role-based pools with ordered failover.

Provider Configuration

Model providers are configured through two files:

  1. config/model-registry.json — Model capabilities and metadata
  2. config/free-first-pools.json — Role-to-model assignments

Supported Providers

opencode-go (Default Primary)

Paid models with zero-retention privacy policy.

Model Roles Capabilities
deepseek-v4-flash Orchestrator, Build, Test, Escalation Coding: 9, Orchestration: 9, Debugging: 8
mimo-v2.5 Review, Reconcile, Trivial Builder Review: 8, Coding: 8, Debugging: 7
deepseek-v4-pro Paid fallback Coding: 9, Review: 9, Debugging: 9

nvidia (Free Tier Fallback)

Model Roles Notes
mistral-large-3-675b Review, Orchestrator fallback Good review quality
mistral-small-4-119b Build, Test fallback Fast execution
llama-3.3-70b Explore Code search
kimi-k2.6 Reconcile Conflict resolution
qwen3-coder-480b Build (currently disabled) Good coding

cerebras (Free Tier Fallback)

Model Roles Notes
gpt-oss-120b Build, Test, Review fallback Fast inference

opencode (Free Tier Fallback)

Model Roles Notes
deepseek-v4-flash-free Free fallback Data may be used for improvement
mimo-v2.5-free Free fallback Promotional free tier
north-mini-code-free Utility Lightweight

openai (Paid Fallback)

Model Roles Notes
gpt-5.6-luna Escalation primary Medium reasoning
gpt-5.5 Escalation fallback Available as backup

Local Ollama 9B

Runs on local GPU (RX 580) for sensitive or offline tasks.

Agent Purpose
local-ollama-explore Read-only exploration
local-ollama-test-fixer Narrow test repair
local-ollama-review Preliminary review
local-ollama-private-worker Sensitive/local-only tasks
local-ollama-builder Small-medium bounded changes

Privacy Classification

Models are classified by data-handling policy:

Classification Description Sensitive Code
paid Zero-retention, trusted provider ✅
trusted-provider-only Contractual data protection ✅
promotional free Limited time, data may be used ❌
data-used-for-improvement Output used for training ❌
data-retained-for-improvement Data retained for improvement ❌
trial-only-logged Trial access, data logged ❌
local-private Runs on local hardware ✅

Adding a Provider

  1. Add model entry to config/model-registry.json
  2. Add model pool entry to config/free-first-pools.json
  3. Add provider timeout to config/free-first-config.json
  4. Optional: Add provider-specific smoke test timeout