M-DNA Forge

Merge 2–5 parent models into hybrid offspring using 12 methods with 3D visualization and presets

M-DNA Forge (08)

M-DNA Forge is ForgeAI’s model merging system. Load multiple parent models, choose a merge strategy, configure layer assignments, and build a hybrid offspring — with 12 merge methods, 5 quick presets, and an interactive 3D visualization.

Concepts

Term	Meaning
Parent	A source model contributing weights (2–5 supported)
Offspring	The resulting merged model
Base model	A reference model required by some methods (TIES, DARE, Task Arithmetic, DeLLa)
Layer assignment	Which parent provides which layer in the offspring

12 Merge Methods

Method	Difficulty	Requires Base	Description
Average	Easy	No	Weighted mean of tensors
SLERP	Easy	No	Spherical interpolation — best for 2-model merges
Passthrough	Easy	No	Direct copy from a single parent
Task Arithmetic	Intermediate	Yes	Add task vectors from finetunes to a base
Frankenmerge	Intermediate	No	Cherry-pick layers from specific parents
DARE	Intermediate	Yes	Drop and rescale delta parameters
TIES	Intermediate	Yes	Trim, elect sign, merge — resolve interference
DeLLa	Advanced	Yes	Density-based layer-level adaptive merging
Component Merge	Advanced	No	Route attention/MLP/norm to different parents
Tensor Surgery	Advanced	No	Per-tensor source mapping from any parent
Parameter Slice	Advanced	No	Dimensional slicing across parents
MoE Conversion	Advanced	No	Convert dense models to Mixture-of-Experts

Quick Presets

Click a preset to instantly configure method and parameters:

Preset	Method	Key Params	Best For
Quick Blend	Average	—	Simple, fast merging
Smooth Merge	SLERP	t = 0.5	Balanced interpolation
Task Tuner	Task Arithmetic	scaling = 1.0	Adding capabilities
Sparse Mix	DARE	density = 0.5	Efficient delta merging
Consensus	TIES	trim = 0.2	Conflict resolution

Difficulty Modes

Mode	Available Methods	Best For
Easy	Average, SLERP, Passthrough	First-time users
Intermediate	+ Frankenmerge, Task Arithmetic, DARE, TIES	Most merges
Advanced	All 12 methods	Expert users

3D Isometric Visualization

Interactive tower view showing all parent models and the resulting offspring:

• Each layer color-coded by source parent
• Hover for detailed tooltips
• Pan, zoom, and reset controls

Layout

Two-column layout with 3D visualization on the left and a tabbed panel on the right:

Files Tab

• Parent model cards (name, params, layers, format, size)
• Load / remove parents
• Set base model (for methods that require one)
• Compatibility report and composition statistics

Layers Tab

• Per-layer parent assignment (click to cycle)
• AUTO: SPLIT — alternating parent assignment
• AUTO: INTERLEAVE — equal chunk assignment
• Layer analysis — specialization badges (Syntactic / Semantic / Reasoning)
• Capability detection and filtering (7 capabilities)

Settings Tab

• Mode selector (Easy / Intermediate / Advanced)
• Presets section with 5 quick-apply cards
• Method grid filtered by current mode
• Method-specific parameter sliders
• Output format (SafeTensors / GGUF), model name, output path
• Batch size slider (1–16 concurrent tensors)

Performance Settings

Setting	Range	Default	Description
Batch Size	1–16	1	Tensors processed concurrently. Higher = faster, more RAM.

Compatibility Checking

With 2+ parents loaded, ForgeAI validates:

• Architecture match
• Hidden dimension match
• Layer count match
• Shared tensor count

Cross-Dimension Merging

When parent models have different hidden dimensions (e.g., 640 vs 1152), ForgeAI can still merge them using a resolution strategy. The strategy adapts tensor shapes before the merge operation runs.

Strategy	Quality	Description
Interpolation	Medium	Resize tensors via linear/nearest-neighbor interpolation to match target shape. Default when dimensions mismatch.
Zero Padding	Medium	Pad smaller tensors with zeros to match the largest parent’s shape.
Truncation	Low	Truncate larger tensors to match the smallest parent’s shape. Loses data.
MoE Routing	High	Convert to Mixture-of-Experts — each parent becomes an expert, bypassing dimension issues entirely.

When a dimension mismatch is detected, Interpolation is auto-selected. Click any strategy in the compatibility panel to switch.

Output Formats

Format	Output
SafeTensors	Directory with model weights + tokenizer/config files
GGUF	Single .gguf file with embedded metadata and tokenizer

Workflow

1. Load parents

   Add 2+ models via the Files tab

2. Check compatibility

   Review the compatibility report

3. Choose preset or method

   Click a quick preset, or manually select mode → method → parameters

4. Assign layers

   For Frankenmerge: set layer assignments (or use AUTO SPLIT/INTERLEAVE)

5. Build

   Click BUILD MERGE, monitor progress in sidebar and status bar

Merge runs in the background — you can navigate freely while it processes. Progress is shown in the sidebar and status bar.