Mixedbread's Toast 1 rewrites search economics

Toast 1 shows that the best agent architecture may be a strong general model paired with a much smaller specialist retriever.

• –Qwen3.6-35B-A3B fine-tuning gives Toast 1 a high-performance base without frontier-model inference costs
• –The agent handles semantic search, grep, reading, filtering, and context management across multimodal enterprise data
• –Retrieval specialization tackles a major bottleneck: frontier models often waste time and tokens finding the right evidence
• –Results push the speed-cost frontier, reaching comparable quality in roughly 8–10 seconds at about $0.08 per query
• –Prime Intellect's reusable harness and RL infrastructure let Mixedbread train on proprietary data without rebuilding its stack