An independent cloud interface for instant model merging
A cloud-based web interface for LLM model fusion — browse merge recipes, generate config.yaml files, and combine Hugging Face models online. Merge Llama, Mistral, Qwen, and more without a GPU, then export to GGUF or Ollama-ready formats.
Get early access + a free Top 50 Merge Recipes PDF. No spam, ever.
2,400+
Recipes indexed
180+
Contributors
50+
Merge methods
Core Platform
Everything the community needs — in one place
Stop jumping between scattered GitHub repos, spreadsheets, and Discord threads. The MergeKit UI brings everything together — create a custom LLM without fine-tuning, right from your browser.
Recipe Registry
A searchable catalog of community-contributed merge recipes with ready-to-use config.yaml files. Filter by SLERP, TIES, DARE, passthrough, or any method — for Llama, Mistral, Qwen, and every supported architecture.
Merge Map Visualizer
Explore the lineage and layer-level composition of any merged model through interactive node graphs. Trace weight flows, identify shared ancestors, and understand merge topology at a glance.
Specialized Leaderboards
Purpose-built rankings for merged models — not just overall scores. Compare GGUF exports, AWQ/GPTQ quantized variants, and full-precision merges across task-specific, safety, and community-voted dimensions.
Model Merging Explained
What is model merging — and why does it matter?
Model merging is a technique that arithmetically combines the weights of two or more independently fine-tuned large language models into a single unified model — without any additional GPU training. Unlike traditional fine-tuning, merging lets you create a custom LLM without fine-tuning budgets. Researchers blend existing expert checkpoints — Llama, Mistral, Qwen, and others — to produce a generalist that inherits complementary capabilities from each parent.
The result is a model that can reason, write code, and follow complex instructions simultaneously — capabilities typically siloed in separate specialist models — while fitting in the same memory budget as a single base model. Community experiments on Hugging Face consistently show merged models occupying the top slots on open leaderboards, often outscoring models three to four times their parameter count.
Training a frontier LLM from scratch costs millions of dollars in compute. Model merging lets the open-source community sidestep most of that cost — you can merge LLMs without a GPU using cloud-based tools or even low-VRAM consumer hardware. A well-constructed merge can outperform any single parent on broad benchmarks, and the result can be exported to GGUF for Ollama, quantized with AWQ or GPTQ, or pushed directly to Hugging Face.
Common merge methods
Spherical Linear Interpolation
Treats model weights as vectors on a high-dimensional sphere and interpolates smoothly between two checkpoints. Produces balanced generalists with minimal capability degradation — ideal for blending a strong reasoner with a strong coder.
Trim, Elect Sign & Merge
Eliminates redundant and conflicting parameter changes before merging. By resolving sign conflicts in weight deltas, TIES dramatically reduces interference when combining models with very different fine-tuning trajectories.
Drop And REscale
Randomly drops a fraction of fine-tuned weight deltas and rescales the remainder. Acts as regularization during the merge, enabling stable combination of many models without catastrophic interference.
Task Arithmetic
Computes a task vector — the delta between a fine-tuned model and its base — for each expert, then adds or subtracts those vectors to steer the merged model's behaviour. Enables surgical capability editing without retraining.
Who is MergeKit for?
MergeKit is built for ML researchers who want to publish reproducible merge experiments, independent developers who need capable open models without fine-tuning budgets, and AI teams who want to track the community's best merged models from a single dashboard. The MergeKit web interface supports every popular architecture — Llama 3, Mistral, Qwen, Phi, and more — and generates config.yaml files you can use directly with the MergeKit GitHub CLI or our cloud GUI. Our recipe registry, interactive visualizer, and purpose-built leaderboards are free to use during beta — no credit card required.
Config Generator
Build your config.yaml instantly
Pick a merge method, add your models, tweak the parameters — get a ready-to-run MergeKit config file in seconds. No installs required.
| Method | Best For |
|---|---|
| SLERP | Merging two high-quality models (e.g. Llama 3 & Mistral) into a balanced generalist. |
| TIES | Combining many task-specific fine-tunes without weight interference or sign conflicts. |
| DARE | Pruning redundant weight deltas for higher efficiency — ideal before a TIES merge. |
| Task Arithmetic | Surgically adding or subtracting capabilities by scaling task vectors over a base model. |
| Passthrough | Stacking layer slices from different models (frankenmerging) to hit a custom parameter count. |
Blend two models via spherical interpolation — ideal for balanced generalists
0 = pure Model A · 0.5 = equal blend · 1 = pure Model B
merge_method: slerpbase_model: meta-llama/Llama-3.1-8B-Instructmodels:- model: meta-llama/Llama-3.1-8B-Instruct- model: mistralai/Mistral-7B-Instruct-v0.3parameters:t: 0.50dtype: bfloat16
mergekit-yaml config.yaml ./output-modelWant to run this in the cloud — no GPU, no CLI? Join the waitlist →
Merge Map Visualizer
See how models are actually built
Every merged model has a hidden family tree. Whether you combined Mistral and Llama via SLERP or built a Qwen passthrough merge, our interactive Merge Maps let you zoom into layer-level detail, trace parameter flows, and spot which source models contributed what.
- Interactive node-graph exploration
- Layer-level weight contribution heatmaps
- Ancestor model lineage tracing
- Exportable SVG & JSON diagrams
| # | Model | Method | Score |
|---|---|---|---|
| 1 | NeuralMerge-70B-v2 | DARE + TIES | 89.4+3.2 |
| 2 | SynthWizard-34B | Linear | 87.1+1.8 |
| 3 | MegaCode-Instruct | SLERP | 86.7+0.5 |
| 4 | AlphaChat-13B-Mix | Task Arithmetic | 85.2-0.3 |
| 5 | OmniMerge-7B | Passthrough | 84.8+2.1 |
Specialized Leaderboards
Rankings built for merged models
Generic leaderboards bury merged models alongside base models. Ours are purpose-built with dimensions that actually matter — safety, code, reasoning, instruction-following, and community-voted categories.
- Task-specific dimension rankings
- Transparent, reproducible eval methodology
- Daily automated re-evaluation
- Community-nominated benchmark categories
Ecosystem
Key Resources
Everything you need to start merging — official tools, benchmarks, and research.
Get early access & the Top 50 Merge Recipes
Join the waitlist for priority access to the registry, visualizer, and leaderboards. As a thank-you, we'll send you a curated PDF of the 50 highest-impact merge recipes — method, config, and benchmark results included.
Free forever during beta · No credit card · Unsubscribe anytime