Make it cheaper?

[TomLucidor]

Considering the advancements of Qwen3-Next and Kimi-Linear, as well as nVidia and whoever else cooked up SLMs to be as capable as LLMs, would it be possible to update this system to be as cost-efficient as possible (with options to pick providers/makers like how OpenRouter is open for everyone) https://github.com/ChaokunHong/MetaScreener?tab=readme-ov-file#step-1--get-your-ai-key-2-minutes

[ChaokunHong]

Thank you for this thoughtful suggestion! You're absolutely right, the landscape has shifted dramatically. Models like Qwen3, Kimi, and various SLMs from NVIDIA and others are now delivering performance competitive with much larger LLMs at a fraction of the cost.
I'm actually already working on MetaScreener 2.0, which addresses exactly these concerns. The key planned improvements include:
Multi-provider support: Integration with OpenRouter alongside direct API access to providers like Qwen (Alibaba Cloud), Moonshot (Kimi), and others — giving users flexibility to choose based on cost, performance, or data privacy needs
Open-source LLM support: Moving beyond commercial-only APIs to include locally deployable models (e.g., via Ollama), which is particularly important for reproducibility in academic research
Cost-optimized ensemble architecture: A Hierarchical Consensus Network where a lightweight SLM handles initial screening and a more capable model adjudicates only uncertain cases — significantly reducing per-article cost without sacrificing sensitivity
Model benchmarking module: So users can evaluate which model best suits their specific review topic before committing to full-scale screening
The current MetaScreener 1.0 was built around OpenAI/Claude APIs as a proof-of-concept, but your point about cost-efficiency is one of the primary motivations behind the redesign.
Stay tuned, and if you'd like to contribute or beta test, feel free to open an issue or reach out directly!