QT 6.5 HOWITZER
DOCS / GETTING STARTED / UNIVERSAL MODELS

Universal Model Orchestration

Project Howitzer decouples prompt engineering and workflow execution from any single model vendor. You have full freedom to orchestrate frontier commercial endpoints and private open-weight GGUF models side-by-side.

1. Frontier Cloud Model Support

Frontier cloud APIs connect over secure TLS with HTTP/2 streaming. All API tokens are stored in the hardware-backed Magazine [MZ] vault and injected directly into request headers at runtime.

PROVIDER FLAGSHIP MODELS STREAMING SPECIAL CAPABILITIES
OpenAI gpt-4o, gpt-4o-mini, o1-preview SSE Native JSON Schema AST enforcement, Vision, Tool Calling
Anthropic claude-3-5-sonnet, claude-3-5-haiku SSE Extended thinking blocks, 200k context window, Artifacts
Google gemini-3-5-flash, gemini-1-5-pro SSE High-frequency B2B execution, 1M+ token context buffer
Mistral mistral-large-2, codestral-2501 SSE Fill-In-the-Middle (FIM) code completions, multi-lingual

2. 1-Click Curated Local Models

Howitzer bundles direct 1-click downloads for pre-quantized GGUF weights optimized for Apple Silicon Metal shaders and Linux Vulkan runtimes.

CODE SPECIALIST 7B / 14B

Qwen 2.5 Coder

Excels at code refactoring, bug triage, unit test generation, and AST manipulation.

  • Quant:Q4_K_M (4.6 GB) / Q8_0 (8.1 GB)
  • RAM:6 GB Unified Memory
  • Offload:100% Metal GPU Offloading
REASONING ENGINE 8B / 14B

DeepSeek R1 Distill

Trained with reinforcement learning for dense step-by-step logic, mathematical problem solving, and formal verification.

  • Quant:Q4_K_M (4.9 GB)
  • RAM:6.8 GB Unified Memory
  • Offload:100% Metal GPU Offloading

3. Loading Custom & External Weights

If you have already downloaded .gguf weights from Hugging Face or compiled private models with MLX, you can load them directly:

  • Open Settings → Inference & Hardware.
  • Click [ + ADD EXTERNAL WEIGHTS ] and browse to your local .gguf path.
  • Set the context length (e.g. 8192, 32768) and GPU layer count (e.g. 99 for full offload).
  • The custom model will appear instantly in the model selection picker across all 5 tactical views.