neuralclones.ai

Don't train from zero.

Get your new architecture as a checkpoint that already knows what a trained model knows. We map your design onto a proven model, keep what transfers, and hand you a head start. You'll know what carries over before you spend a dollar on compute.

In one call, you'll learn which of your changes transfer and what a pilot would take.

We beat the benchmark average of a model that costs up to $26,000 to train, for under $1,200 of compute.

More than 21× cheaper. Over 70× against what Gensyn's run actually used.

Our Qwen3.5 hybrid against Gensyn's Open-1B, on Open-1B's own ten-task suite. Read the case study →

The problem

Change the architecture, and your checkpoints stop fitting.

Faster attention, smaller memory, better hardware fit: new designs arrive every month. But change a model's architecture and its trained weights no longer load. The knowledge you paid for gets left behind.

So new designs often pay to learn it all over again.

36–400B+

Tokens Open-1B had trained on when it first matched our Qwen3.5 hybrid, by test. On three tests it never did, even at its 400B finish.

0.37B

Tokens of training our Qwen3.5 hybrid used, starting from a published Qwen3.5 0.8B.

Qwen3.5 0.8B mostly keeps a running summary of the text, while Open-1B looks back at the text directly with attention in every layer. Our Qwen3.5 hybrid combines both: Open-1B-style attention in 21 of its 24 layers, with Qwen's first three layers and two of its own parts kept where they matter most. See the case study.

The solution

Build on training that's already been done.

Build your new model on top of the training an existing model has already done. We integrate the advantages of your technology into a model that's already trained, keeping training worth tens of billions, hundreds of billions, even trillions of tokens.

How it works

Bring your ideal architecture. We hand you a head start.

  1. Send us your ideal architecture

    The design you want to train, as a configuration or as code. That's all we need: we recommend the trained model to start from.

  2. We scope the transformation

    We map every change from the source model to your design and the effect of each one: what carries across exactly, what has to be fitted, and what it costs. Before any compute is spent.

  3. We settle the final architecture

    We determine the final design with you. We keep every change that's worth its cost and adjust any that isn't.

  4. We convert and fit. You train.

    Every part is rebuilt to the final design and trained to do what the original did. You get a head start to train from, with a measured report on its benchmarks and long-range memory.

Case study

A Qwen3.5 / Open-1B hybrid that outscores Open-1B

Qwen3.5 0.8B and Gensyn's Open-1B combined into one architecture: Qwen's first three layers and two of its own parts kept, the rest rebuilt as Open-1B's attention.

52.8on Open-1B's ten-task benchmark, against Open-1B's 50.3
7 of 10tasks where it beats Open-1B
~370Mtokens of training, against Open-1B's 400B
Read the full case study

Our promise

Every checkpoint comes with the truth about it.

Next step

Training a new architecture? Start ahead.

In one call, you'll learn which of your changes transfer and what a pilot would take.

Discuss your architecture

Figures come from our own experiments. The $26,000 is our estimate for training Open-1B at normal speed; Gensyn's actual run cost $86,000 to $128,000 at today's prices. Token comparisons start from a published trained model whose original training is not counted, and our training tokens include a teacher model running alongside. Qwen is published by the Qwen team and Open-1B by Gensyn; Neural Clones is independent of both.