LEVI
run AlphaEvolve on your Claude Code/Codex for dirt cheap
Last verified:
What is LEVI?
LEVI (LLM Evolution through Voronoi Initialization) is an open-source harness-first framework for LLM-guided evolutionary search over code and prompts. It is designed to make LLM-guided optimization (like AlphaEvolve, GEPA, ShinkaEvolve) much more accessible and affordable without sacrificing results. The framework points at a scoring function and a budget, then evolves the artifact automatically using API calls to LLMs.
LEVI achieves the highest scores across all competing frameworks on the ADRS benchmark with 1.5–6.7x cost savings. It beats OpenEvolve/GEPA at 1/6th the cost and beats AlphaEvolve's Circle Packing while mostly using a 30B model. The framework uses two core ideas: stratified model allocation (larger models for rare paradigm shifts, smaller cheaper models for majority of mutations) and improved diversity maintenance (CVT-MAP-Elites archive with Voronoi initialization).
The tool is for researchers and practitioners who want to optimize code or prompts using evolutionary search but cannot afford large budgets for frontier-scale models. It enables the average researcher to use LLM-guided optimization methods without burning through large budgets, making this promising direction more accessible.
Key benchmark results include Spot Single-Reg (51.7 vs GEPA 51.4, 6.7x cheaper), Spot Multi-Reg (72.4 vs OpenEvolve 66.7, 5.6x cheaper), LLM-SQL (78.3 vs OpenEvolve 72.5, 4.4x cheaper), Cloudcast (100.0 vs GEPA 96.6, 3.3x cheaper), Prism (87.4 tied, 3.3x cheaper), EPLB (74.6 vs GEPA 70.2, 3.3x cheaper), and Txn Scheduling (71.1 vs OpenEvolve 70.0, 1.5x cheaper).
LEVI pricing
Pricing model: Freemium
LEVI is completely free as an open-source framework available on GitHub. There is no paid tier or subscription. The cost savings mentioned ($4.50 per problem on most tasks vs $15–$30 for baselines) refer to the LLM API costs users pay when running LEVI, not fees to LEVI itself. Users only pay for the LLM API calls they make (approximately 90% to lightweight models like Qwen3-30B-A3B and MiMo-v2-Flash, 10% to larger models like Gemini Flash 3 for paradigm shifts).
LEVI pros
- Achieves highest scores on all ADRS benchmark problems
- 1.5–6.7x cost savings compared to competing frameworks
- Beats OpenEvolve/GEPA at 1/6th the cost
- Beats AlphaEvolve's Circle Packing using Qwen 30B for 95%+ mutations
- Open-source framework available on GitHub
- Stratified model allocation matches model capacity to task demand
- 90% of LLM calls routed to lightweight cheaper models
- CVT-MAP-Elites archive with Voronoi initialization maintains diversity
- Combines structural and functional diversity dimensions
- Sustains exploration well past when baselines converge
- Reaches near-peak performance 12x faster than OpenEvolve
- Perfect 100.0 score on Cloudcast problem
- $4.50 per problem on most tasks vs $15–$30 for baselines
- Data-driven initialization with noise prevents overfitting
- Works with smaller models like Qwen 30B effectively
- Strong sample efficiency with fewer evaluations needed
- Unified fingerprint space with AST-based structural features
LEVI cons
- Still in development, not fully released yet
- Requires user to provide scoring function and seed program
- Depends on access to multiple LLM APIs
- Computational overhead from maintaining CVT-MAP-Elites archive
- May need tuning for specific problem domains
- 90% cheap model usage may miss some paradigm shifts
- Limited to code and prompt optimization
- Requires technical expertise to set up scoring functions
- Archive geometry may overfit to early seed families without noise
Frequently asked questions about LEVI
What is LEVI and what does it do?
LEVI (LLM Evolution through Voronoi Initialization) is an LLM-guided evolutionary framework for code and prompts. Point it at a scoring function and a budget—LEVI evolves the artifact for you using API calls to LLMs. It beats OpenEvolve/GEPA at 1/6th the cost and achieves the highest scores on the ADRS benchmark.
How does LEVI's stratified model allocation work?
LEVI allocates larger, more capable LLMs for rare but creative paradigm shifts (about 10% of calls) and smaller, cheaper LLMs for the majority of mutations (about 90%). Larger models suggest creative working code with higher chance of moving evolution forward, while smaller models handle local refinements. A Qwen 30B is 40x cheaper than GPT 5.4, making this trade-off economically reasonable.
What is LEVI's diversity maintenance mechanism?
LEVI uses CVT-MAP-Elites with Voronoi initialization, combining AlphaEvolve's MAP-Elites approach (maintaining solutions along a grid of diversity dimensions) and GEPA's Pareto frontier (different algorithms have different trade-offs). It uses both structural behaviors (number of math operators, cyclomatic complexity) and functional behaviors (Pareto-like trade-offs) as dimensions in a unified fingerprint space.
How does LEVI compare to OpenEvolve and GEPA on benchmarks?
LEVI achieves the highest score on every ADRS benchmark problem where improvement is possible, with an average of 76.5 compared to 71.9 for GEPA (next-best), a +4.6 point improvement. Specific results: Spot Single-Reg 51.7 vs 51.4 (6.7x cheaper), Spot Multi-Reg 72.4 vs 66.7 (5.6x cheaper), LLM-SQL 78.3 vs 72.5 (4.4x cheaper), Cloudcast 100.0 vs 96.6 (3.3x cheaper).
What LLMs does LEVI use?
LEVI routes approximately 90% of LLM calls to lightweight models (Qwen3-30B-A3B and MiMo-v2-Flash) and the remaining 10% to larger models (Gemini Flash 3) for paradigm shifts. In controlled comparisons, LEVI works effectively with just a locally-served Qwen3-30B-A3B model.
How much does it cost to run LEVI?
LEVI uses $4.50 per problem on most tasks (Transaction Scheduling: $13) versus baselines' $15 to $30. The per-generation cost drops by roughly 10x compared to baselines using GPT-5 or Gemini-3.0-Pro for every call, enabling more generations at lower total spend.
Is LEVI open source?
Yes, LEVI is available as an open-source framework at github.com/ttanv/levi. It will be fully open-sourced on GitHub soon. Users can point it at a scoring function and a seed program and it runs until the budget is spent.
What problems can LEVI optimize?
LEVI is designed for code and prompt optimization. On the ADRS benchmark, it has been evaluated on real-world systems problems spanning cloud scheduling (Cloudcast), load balancing (Spot Single-Reg, Spot Multi-Reg), SQL optimization (LLM-SQL), transaction scheduling (Txn Scheduling), mathematical constructions (Prism, EPLB), and simulation optimization (Can't Be Late).
How does LEVI initialize its archive?
LEVI uses data-driven initialization with noise. It generates fewer than 10 structurally distinct seed programs, expands them into variants, fingerprints them all, then adds Gaussian noise before fitting centroids. The seeds anchor the tessellation in regions where viable programs occupy, while noise broadens each family's footprint to accept innovations between or outside initial seed families.
What makes LEVI different from other evolutionary frameworks?
LEVI takes a harness-first approach instead of building around the assumption of strong frontier models. It invests in the search architecture (stratified model allocation and improved diversity maintenance) rather than relying on larger models. This reduces the burden on the model, making it possible to get strong results with cheaper models and smaller budgets while maintaining or improving raw results.