Show HN: We trained an MoE LLM built for developer tasks (interfaze.ai)

Interfaze-beta is a new Mixture-of-Experts (MoE) large language model suite built and tuned specifically for developer workflows. Rather than one huge monolith, Interfaze routes tasks to a collection of smaller specialist models on custom infrastructure, supports multimodal inputs (audio, images, PDF, CSV, JSON, etc.), and integrates tools like web search and code-run sandboxes to validate outputs and reduce hallucinations. It’s OpenAI Chat API–compatible (swap base URL/keys), supports structured JSON or plain-text outputs, and exposes configurable safety guardrails for text and images—making it easy to drop into existing SDKs and pipelines. Technically, Interfaze uses native expert routing and optional activation of a stronger reasoning model for harder problems, trading top-tier general knowledge for a developer-optimized balance of speed, cost, and reliability. Reported benchmarks put it near or above current state-of-the-art on many developer-relevant tasks: ~90% on ChartQA/AI2D, top AIME-2025 score at 90%, strong LiveCodeBench v5 coding results, and narrowly behind Gemini-2.5-Pro on high-end reasoning benchmarks. Practical implications: faster, cheaper, and more controllable developer automation (OCR, scraping, classification, multimodal understanding, multi-turn reasoning) with built-in verification and safety—appealing for teams prioritizing reliable, structured outputs over broad encyclopedic performance.