| name: Neuro-Orchestrator-8B | |
| merge_method: ties | |
| base_model: miromind-ai/MiroThinker-v1.0-8B | |
| parameters: | |
| normalize: true | |
| int8_mask: true | |
| dtype: bfloat16 | |
| tokenizer_source: union # Critical: Merges special tokens from all models | |
| models: | |
| # Base Model: MiroThinker (The Body/Executioner) | |
| # We use this as the base so the resulting model defaults to its | |
| # superior 256k context and tool-use templates. | |
| - model: miromind-ai/MiroThinker-v1.0-8B | |
| parameters: | |
| density: 1.0 | |
| weight: 1.0 | |
| # Model 2: Nvidia Nemotron Orchestrator (The Manager) | |
| # We want high density (coverage) but lower weight to guide | |
| # logic without breaking the tool syntax. | |
| - model: nvidia/Nemotron-Orchestrator-8B | |
| parameters: | |
| density: 0.8 | |
| weight: 0.4 | |
| # Model 3: Kwaipilot HiPO (The Controller) | |
| # We mainly want its 'gating' logic (when to think). | |
| # A lower density focuses on the most critical policy changes. | |
| - model: Kwaipilot/HiPO-8B | |
| parameters: | |
| density: 0.5 | |
| weight: 0.3 | |