What is ShinkaEvolve?
ShinkaEvolve is an open-source framework from Sakana AI for improving code by evolution. You supply a starting program and an evaluator that scores it; an ensemble of LLMs proposes edits, the evaluator scores each candidate, and the best survive across generations and separate islands of programs. A bandit picks which model to call for each mutation, and novelty filtering avoids paying to evaluate near-duplicates. It is inspired by Google DeepMind's AlphaEvolve, which is not open source, and by Sakana's own AI Scientist and Darwin Gödel Machine.
Sakana's launch results include a new best 26-circle packing found in 150 samples, an AIME agent scaffold evolved in 75 generations, a 2.3% average gain on ALE-Bench programming tasks, and a mixture-of-experts load-balancing loss that beat DeepSeek's after 30 generations. The paper was accepted at ICLR 2026, and the framework supported the team that won the ICFP 2025 Programming Contest.
The code is Apache 2.0 and on PyPI as shinka-evolve; the latest release is v0.0.7 from 2 June 2026 and the last commit was on 5 October 2026. It has about 1,480 GitHub stars.
Is ShinkaEvolve worth it?
ShinkaEvolve is the strongest open-source choice for LLM-driven program search, and its selling point is sample efficiency. Earlier evolutionary approaches often needed thousands of LLM calls per problem; Sakana's headline result is a new best 26-circle packing in 150 samples, achieved through rejection sampling of near-duplicate candidates, a bandit that shifts calls to whichever model is improving, and islands that share an archive of good solutions. Those are Sakana's own experiments, but the ICLR 2026 acceptance and the ICFP 2025 contest win give them more outside support than most framework claims receive.
It fits a narrow and valuable shape of problem: a program with an automatic verifier and a number to improve. The shipped examples cover circle packing, a 2048 policy, Julia, Fortran and Wolfram solvers, and Verilog designs scored for power, performance and area. It is not a general coding agent and does nothing useful without an evaluator you trust; a weak evaluator will be gamed.
The project has matured quickly in 2026: a unified ShinkaEvolveRunner in March (which replaced the old runner classes, so 2025 code needs changes), PyPI packaging, a documentation site in April, agent skills for Claude Code and Codex, and in May Headless CLI support for routing mutations through a Claude or ChatGPT subscription instead of API billing. Versions are still 0.0.x, so expect further breaking changes.
Safety is left to you. Every candidate is LLM-written code that your evaluator executes, by default directly on the host or on a Slurm cluster, and the getting-started guide says nothing about sandboxing. Slurm Docker jobs give isolation on a cluster; on a workstation, run it in a container. The default embedding model is OpenAI's text-embedding-3-small, so code embeddings leave your machine unless you point it at a local endpoint.
Where it earns its place
An operations-research engineer with a scheduling heuristic
Improve the heuristic's score on a fixed benchmark set
Wrap the benchmark as evaluate.py, set max_api_costs to cap spend, and run a few dozen generations. Sakana's ALE-Bench runs averaged a 2.3% gain over strong baselines.
A hardware team optimising Verilog blocks
Search for RTL that lowers power and area under a fixed spec
Start from the RTLLM example, which needs iverilog, yosys and OpenSTA, and let candidates be scored on power, performance and area. Run evaluations in Slurm Docker jobs to contain generated code.
An ML researcher with a Claude subscription and no API budget
Evolve a training-loss variant without per-token billing
Use headless/claude model strings so mutations run through the subscription, as Sakana's sine_approx_headless example shows. Sakana's MoE loss result came after 30 generations.
Before you commit
- Setup
- pip install shinka-evolve (Python 3.10 or later, 3.11 recommended), add provider keys to .env, then shinka_launch variant=circle_packing_example; the circle-packing task is the documented first run.
- Integrations
- OpenAI, Anthropic, Gemini and OpenRouter models, any local OpenAI-compatible endpoint (such as a local Ollama server), Headless CLI for Claude and Codex subscriptions, Slurm (Conda or Docker), a WebUI, Weights & Biases, and agent skills for Claude Code and Codex.
- Data and privacy
- Code and prompts go to the LLM providers you configure, and to OpenAI for embeddings by default. Runs are stored in a local SQLite database. W&B logging, if enabled, uploads an allowlisted set of metrics rather than code. Apache 2.0 places no limits on commercial use.
- Cost at scale
- Cost is your LLM calls plus evaluator compute; Sakana publishes no dollar figures. The async runner, which Sakana says is 5 to 10 times faster, raises throughput and spend together, so set max_api_costs on long runs.
- Limits
- Only useful where an automatic evaluator exists. No sandbox for evaluated code on local runs. Version 0.0.7, with 14 open issues and pull requests and API changes during 2026.
What's good
- Sample-efficient: new best 26-circle packing in 150 samples, per Sakana
- Model ensemble with bandit selection and cost-aware weighting
- Runs locally or on Slurm, with a WebUI and W&B logging
- Hard budget cap in USD via max_api_costs
- Apache 2.0, on PyPI, ICLR 2026 paper
What's not
- Executes LLM-written code with no built-in sandbox
- Needs a reliable automatic evaluator to be useful
- Pre-1.0; the March 2026 runner refactor broke older code
- Embeddings default to OpenAI unless configured otherwise
- Fetches live model prices from models.dev at start unless set offline
Best for
Researchers and performance engineers who have a scored program, such as a heuristic, kernel or solver, and want LLMs to search for better versions within a fixed budget.
ShinkaEvolve alternatives
ShinkaEvolve: common questions
- Is ShinkaEvolve worth it?
- For anyone with a program and a score to push up, this is the open alternative to AlphaEvolve to start with, provided they sandbox the code it writes. We rate it 4.1 out of 5.
- How much does ShinkaEvolve cost?
- Free under Apache 2.0. You pay for the LLM calls (or a Claude or ChatGPT subscription through Headless) and the compute that runs your evaluator.
- What are the downsides of ShinkaEvolve?
- Executes LLM-written code with no built-in sandbox; needs a reliable automatic evaluator to be useful; pre-1.0; the March 2026 runner refactor broke older code; embeddings default to OpenAI unless configured otherwise; fetches live model prices from models.dev at start unless set offline.
- Who is ShinkaEvolve best for?
- Researchers and performance engineers who have a scored program, such as a heuristic, kernel or solver, and want LLMs to search for better versions within a fixed budget.
- How much does ShinkaEvolve cost for a team or in production?
- Cost is your LLM calls plus evaluator compute; Sakana publishes no dollar figures. The async runner, which Sakana says is 5 to 10 times faster, raises throughput and spend together, so set max_api_costs on long runs.
- Is ShinkaEvolve safe for company data?
- Code and prompts go to the LLM providers you configure, and to OpenAI for embeddings by default. Runs are stored in a local SQLite database. W&B logging, if enabled, uploads an allowlisted set of metrics rather than code. Apache 2.0 places no limits on commercial use.
- What are the best ShinkaEvolve alternatives?
- The closest alternatives we have reviewed are The AI Scientist (v1, v2 and v3), TreeQuest, OpenHands and mini-swe-agent.
Answers checked Oct 26, 2026.
Make ShinkaEvolve? Show this rating on your site
