AIWG Supported Models

Canonical reference for models tested with AIWG. Cloud models are built-in; local models are validated by the AIWG team or community contributors.

---

Cloud Providers (Built-in)

These models are fully supported and configured automatically when using AIWG with cloud providers.

Anthropic

Model	AIWG Tier	Context	Tool Use	Status
claude-opus-4-6	opus	200K	Yes	Fully supported
claude-sonnet-4-6	sonnet	200K	Yes	Fully supported
claude-haiku-4-5-20251001	haiku	200K	Yes	Fully supported

OpenAI

Model	AIWG Tier	Context	Tool Use	Status
gpt-5.4	opus	256K	Yes	Fully supported
gpt-5.3-codex	sonnet	256K	Yes	Fully supported
gpt-5.1-codex-mini	haiku	192K	Yes	Fully supported

Google

Model	AIWG Tier	Context	Tool Use	Status
gemini-2.5-pro	opus	1M	Yes	Fully supported
gemini-2.5-flash	sonnet	1M	Yes	Fully supported

---

Local Models — Validated

Models tested with AIWG by the team or community, with confirmed compatibility results.

Model	Size	Backend	AIWG Tier	Tool Use	Coding	Reasoning	SDLC Workflows	Tested By	Notes
qwen3.5:9b	9B	Ollama	sonnet	Yes	Good	Good	Partial	Community	Strong coding, vision support (256K context), reasoning degrades in long chains
hermes-3-llama-3.1:8b	8B	Ollama	haiku	Yes	Good	Moderate	Limited	Community	Good instruction following, limited multi-step planning
codellama:34b	34B	Ollama	sonnet	No	Excellent	Moderate	Limited	AIWG team	Strong code gen, no native tool use support
qwen2.5-coder:14b	14B	Ollama	sonnet	Yes	Excellent	Moderate	Partial	AIWG team	Best coding quality per VRAM of validated local models
llama3.3:70b	70B	Ollama/vLLM	opus	Partial	Good	Good	Partial	AIWG team	Requires 48GB+ VRAM; closest to cloud quality
llama3.2:3b	3B	Ollama	haiku	No	Basic	Basic	No	AIWG team	Fast summaries and formatting only
llama3.1:8b	8B	Ollama	haiku	No	Moderate	Moderate	No	AIWG team	General purpose, budget tier
codellama:13b	13B	Ollama	haiku	No	Good	Basic	No	AIWG team	Budget code generation

Capability Legend

• Yes — Works reliably in testing
• Partial — Works with limitations (see Notes)
• No — Not supported or not functional
• Good/Moderate/Basic/Excellent — Relative quality rating

Status Definitions

Status	Meaning
Fully supported	Built into AIWG, tested with every release
AIWG team tested	Validated by maintainers, not regression-tested
Community tested	Reported working by community contributor (see notes for details)
Expected compatible	Should work based on architecture, not formally tested

---

Local Models — Expected Compatible (Untested)

Models that should work with AIWG based on architecture and capabilities, but haven't been formally validated.

Model	Size	Expected Tier	Notes
deepseek-coder-v2:16b	16B	sonnet	MoE architecture, strong coding benchmarks
mistral:7b	7B	haiku	Good instruction following
mixtral:8x7b	47B (MoE)	sonnet	MoE architecture, good general capability
phi-4:14b	14B	sonnet	Microsoft, strong reasoning for size
gemma2:27b	27B	sonnet	Google, good structured output
command-r:35b	35B	sonnet	Cohere, tool use support

---

Tier Mapping Guide

When configuring local models, map them to AIWG tiers based on their demonstrated capabilities:

AIWG Tier	Expected Capability	Typical Model Size
opus	Complex reasoning, architecture review, multi-step planning	70B+
sonnet	Code generation, debugging, structured output, moderate reasoning	13B–34B
haiku	Summaries, formatting, simple transformations, fast tasks	3B–8B

Example models.json for Local Setup

{
  "openai": {
    "reasoning": {
      "model": "qwen3.5:9b",
      "description": "Local Qwen 3.5 for coding and reasoning"
    },
    "coding": {
      "model": "qwen2.5-coder:14b",
      "description": "Local Qwen 2.5 Coder for implementation"
    },
    "efficiency": {
      "model": "llama3.2:3b",
      "description": "Local 3B for quick tasks"
    }
  },
  "shorthand": {
    "opus": "qwen3.5:9b",
    "sonnet": "qwen2.5-coder:14b",
    "haiku": "llama3.2:3b"
  }
}

---

Contributing Model Test Results

See Community Model Testing Guide for how to submit your own model test results.

---

See Also

• Local Model Deployment Guide — Setup instructions for Ollama, llama.cpp, vLLM
• Hybrid Architectures — Routing between local and cloud models
• Community Model Testing — How to contribute test results