regolo.ai

The debate over whether open weights can power mission-critical software has effectively ended. Two major ecosystem releases last week marked the definitive inflection point: Mistral Large 4 shipped under open weights on October 6, and GLM-5.3 entered general availability on AWS Bedrock on October 5. When frontier weights deploy directly into hyperscaler catalogs, technical leadershipContinue reading "Hybrid routing is the default: why managed open weights beat self-hosting and closed APIs" The post Hybrid routing is the default: why managed open weights beat self-hosting and closed APIs appeared first on regolo.ai.

A baseline RAG pipeline scores 45.6 percent exact match on HotpotQA and 10.8 percent joint score, because the second document a question depends on never reaches the context window (Yang and colleagues, 2018). In a nutshell. An agentic knowledge graph links entities to the text chunks that mention them, so retrieval follows relational edges insteadContinue reading "Build an agentic knowledge graph that cites every claim" The post Build an agentic knowledge graph that cites every claim appeared first on regolo.ai.

Coding agent benchmarks obsess over model reasoning capacity, but in production harnesses, over 65 percent of your token spend is wasted re-transmitting identical compiler logs, bloated tool outputs, and redundant confirmation turns. When teams deploy autonomous agents across enterprise repositories, developer excitement quickly turns into billing shock as multi-turn debugging sessions ingest hundreds of thousandsContinue reading "Cut coding agent token costs with Sol-Pi and Regolo" The post Cut coding agent token costs with Sol-Pi and Regolo appeared first on regolo.ai.

Agentic rag is an advanced retrieval architecture where an autonomous agent coordinates retrieval actions by evaluating evidence completeness before generating a final answer. Retrieval-augmented generation supplies outside data to a language model, but traditional pipelines execute only one static search pass that cannot recover from initial retrieval errors. Agentic workflows resolve this limitation by introducingContinue reading "Agentic RAG Tutorial: build retrieval that reasons" The post Agentic RAG Tutorial: build retrieval that reasons appeared first on regolo.ai.

Bonsai 2 and Qwen3.8-27B belong to the same approximately 27-billion-parameter family. Bonsai compresses the Qwen parent into ternary weights: its developer reports a capabilities average of 84.78/100, versus 86.32/100 for the original Qwen3.8-27B FP16 reference, across the same 14 thinking-mode benchmarks. On Regolo, we offer Qwen3.8-27B through our managed API at €0.50 per million inputContinue reading "Bonsai 2 vs Qwen3.8-27B: enterprise benchmarks and deployment choices" The post Bonsai 2 vs Qwen3.8-27B: enterprise benchmarks and deployment choices appeared first on regolo.ai.