MagenticBrain-GGUF
GGUF quantization of microsoft/MagenticBrain (14B) for use with llama.cpp and compatible runners (Lemonade, LM Studio, Ollama, etc.).
Available quants
| File | Quant | Size | Notes |
|---|---|---|---|
MB-Q6_K.gguf | Q6_K | ~11.5 GB | Recommended — near-lossless; preserves the model's structured tool-call / JSON fidelity |
MB-Q4_K_M.gguf | Q4_K_M | ~9 GB | Fallback for tight VRAM budgets |
Converted with convert_hf_to_gguf.py (BF16 intermediate) and quantized with llama-quantize from llama.cpp. Standard k-quants, no imatrix calibration.
About the model
MagenticBrain is Microsoft Research AI Frontiers' 14B orchestration model, SFT'd from Qwen3-14B on agentic data — function calling (APIGen-MT, ToolACE, xLAM), 250+ synthetic MCP environments, file-system and terminal trajectories, and sub-agent delegation traces (handoffs to Fara1.5-9B) — followed by an RL stage on terminal tasks. It plans multi-step tasks, calls declared tools via structured JSON (never inventing new ones), coordinates sub-agents, and follows a submit-to-terminate protocol.
This is an orchestration-first model, not a general-purpose chat model. It is co-designed with, and most thoroughly evaluated in, Magentic-Lite (Magentic-UI v0.2).
Recommended settings
- Context: up to 32K
- Thinking: disabled by default (
enable_thinking: false) — keep it off; verbose reasoning degrades long agentic trajectories - Tool calling: schemas are injected by the harness; the model expects to select only from declared tools
Usage
llama.cpp:
llama-server -m MB-Q6_K.gguf -ngl 99 -c 32768
Lemonade / LM Studio / Ollama: register the GGUF file directly or pull this repo through the app's model management.
License & credit
MIT — same as the base model. All credit to Microsoft Research AI Frontiers. This repo is a community quantization and is not affiliated with or endorsed by Microsoft. See the base model card for full details.