More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Model Merging: Combine Fine-Tunes Without Retraining

    1

    Combine fine-tuned models into one without retraining — decide when merging beats training, pick the method by goal, validate the merge.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Model Merging: Combine Fine-Tunes Without Retraining

    Model Merging: Combine Fine-Tunes Without Retraining

    Example session with this skill installed

    We have a math fine-tune and a code fine-tune on the same base model. Merge them into one assistant and prove it did not get worse at either skill.

    • Read your context and instructions
    • Compiled the model merging

    Decision first: both fine-tunes share the same base model, the goal is one assistant that does both, and retraining on combined data would take a week — merging is the right call, and routing to two models was ruled out because the deployment only serves one. Method: linear interpolation was tried first as the baseline, then spherical interpolation for the final merge — the goal is blending two capabilities, not resolving a conflict, so the simpler methods fit. DARE and TIES were noted as the reserve if the blend showed interference. Run: mergekit config written with both models at equal weight, run on CPU in 11 minutes, output saved as a new model with its own version and a record of the exact inputs and weights. The config is version-controlled, so the merge is reproducible. Validation: the merged model was evaluated on the math benchmark and the code benchmark the two parents were measured on — math 71.4 (parent 72.1), code 68.9 (parent 69.5). Both slightly below the specialists, both far above the base model, and the combined assistant now answers mixed questions the parents could not. Regression check: ran the base model's general benchmark too, and the merge held within 1 point — no catastrophic forgetting hiding behind the two good scores. A spot check of 20 mixed questions found 17 good answers, 2 partial, 1 wrong, and the wrong one was a math question with a tricky setup the math parent also missed. Verdict: ship the merge. It keeps 98% of each specialist's skill in one model, the validation is on the record, and the fallback is clear — if production metrics slip, the two specialists are still in the registry.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Two fine-tunes of the same base model each learned something, and merging blends their weights into one model that sometimes keeps both skills — in minutes, on CPU, for the cost of electricity. The discipline is knowing when it works, which method fits the goal, and how to prove the merge did not break anything. This skill covers the merge-train-or-route decision, method selection by goal with mergekit, a full run end to end, validation that treats the merge as guilty until proven innocent, and a worked example merging a math assistant with a code assistant. The warning is plain: a merge that looks clever on paper and regresses in production is worse than no merge at all.

    Merging only works between fine-tunes of the same base model — different bases cannot be merged this way.
    A merge keeps most of each skill, not all of it; the skill shows how to measure the loss, not how to eliminate it.
    Validation on two benchmarks is not proof against every regression — production metrics are the final check.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 3 days ago

    Permissions

    Terminal / Shell
    Listed3 days ago

    What's inside

    Frequently Asked Questions

    No. Merging blends weights that share a coordinate system — the same base model. Different bases have different weight spaces, and the merge is meaningless.

    Probably interference: the two fine-tunes pushed the same weights in opposite directions. Try TIES or DARE, which resolve conflicts, or adjust the weights — and validate on both benchmarks before judging.

    When the goal needs new capability neither parent has, or when merging keeps losing too much. Merging blends what exists; training creates what does not.

    New in LLM Ops