News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
EvoLLM-JP

EvoLLM-JP

Experimental Japanese language model created by merging open-source models.

Sakana AI released EvoLLM-JP as an experimental prototype in 2024. It combines three existing seven-billion parameter models using an evolutionary algorithm to create a Japanese language model with mathematical reasoning skills. The project demonstrates automated model composition without extensive retraining.

Share on WhatsApp

  • Created by merging Shisa Gamma, WizardMath, and Abel models.
  • Uses evolutionary optimization to combine model weights automatically.
  • Achieved strong results on Japanese benchmarks despite small size.
  • Released under Microsoft Research License Terms for research use.

What it is

EvoLLM-JP is an experimental general-purpose language model designed for Japanese text. Sakana AI developed it using a technique called Evolutionary Model Merge. This method automatically discovers effective ways to combine diverse open-source models. The resulting system aims to harness collective intelligence from existing weights without requiring massive new training data or computational resources. It serves as a research prototype rather than a commercial product.

How it came about

The model emerged from research into automated model composition. Takuya Akiba and colleagues at Sakana AI published their findings in March 2024. Their paper described how evolutionary optimization could merge different models in parameter and data flow spaces. The work was later published in Nature Machine Intelligence in January 2025.

Inside the model

This autoregressive transformer was built by merging three specific seven-billion parameter models. The components included Shisa Gamma 7B v1, WizardMath 7B V1.1, and Abel 7B 002. The evolutionary algorithm optimized the combination in both parameter space and data flow space. This approach allowed cross-domain merging, bringing mathematical reasoning capabilities into a Japanese language framework. No extensive additional training data was used for this specific merge.

What it can do

The model focuses on Japanese language tasks and mathematical reasoning. The creators claim it achieved state-of-the-art performance on several established Japanese benchmarks. It reportedly surpassed models with significantly more parameters in these tests. The underlying method also generated a culturally aware visual language model, though EvoLLM-JP itself is text-based. Performance outcomes are not guaranteed as it remains an experimental prototype.

How to use it

Weights are available on Hugging Face under the Microsoft Research License Terms. This license applies due to the inclusion of the WizardMath component. Sakana AI states the model is for research and development purposes only. It is not intended for commercial use or deployment in mission-critical environments. Users must accept all risks associated with its experimental nature.

Safety and testing

The developers warn that performance and outcomes are not guaranteed. They advise users to understand the risks fully before use. The model is described as uncensored in its system prompt example. Sakana AI disclaims liability for any damages arising from its use. It is explicitly marked as an experimental prototype not suitable for critical applications.

Reception and debate

The work introduced a new paradigm for automated model composition. It contributed new models back to the open-source community. The approach demonstrated that effective combinations could be found without human intuition or domain knowledge. The publication in Nature Machine Intelligence highlighted its significance in efficient foundation model development.

Why it matters

This project paved the way for exploring alternative approaches to building foundation models. It showed that evolutionary algorithms could optimize model merging recipes effectively. The success suggested that collective intelligence from open-source models could be harnessed efficiently. It reduced the need for substantial computational resources typically required for developing capable language models.

Questions people ask about EvoLLM-JP

What is EvoLLM-JP?

It is an experimental Japanese language model created by Sakana AI. The model was built by merging three existing open-source models using an evolutionary algorithm. It has seven billion parameters and focuses on general Japanese tasks and math reasoning.

How was this model trained?

It was not trained from scratch. Instead, it was created by merging the weights of Shisa Gamma 7B, WizardMath 7B, and Abel 7B. An evolutionary algorithm optimized the combination to maximize performance without extensive new data or compute.

Can I use this model commercially?

No. Sakana AI states it is for research and development purposes only. It is not intended for commercial use or mission-critical deployment. The license is the Microsoft Research License Terms, and users assume all risks.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on EvoLLM-JP yet, so the facts on this page come from the sources listed above: huggingface.co/SakanaAI/EvoLLM-JP-v1-7B, arxiv.org/abs/2403.13187. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.