InternScience Releases Sage-A1.5 35B Mixture-of-Experts Agentic Model
InternScience has released Sage-A1.5, a 35B Mixture-of-Experts agentic model fine-tuned from Qwen3.6-35B-A3B for long-horizon scientific problem solving. The model introduces enhanced capabilities including agentic orchestration, built-in self-verification, and open scientific discovery support. This release provides the open-source community with an accessible foundation for complex scientific exploration and multi-step reasoning tasks. By combining MoE efficiency with advanced agentic behaviors like parallel subagent delegation, it bridges the gap between raw LLM inference and autonomous research workflows. Sage-A1.5 features an opt-in Verification Mode that allows the model to pause mid-chain-of-thought to self-review its trajectory without external critics. It is natively equipped with function calling and tool integration, and GGUF quantized versions (Q4_K_M and Q8_0) are made available for local hardware deployment.
## BACKGROUND
Mixture-of-Experts (MoE) is a neural network architecture where only a subset of pathways or 'experts' are activated for each token, allowing models to scale up total parameters while keeping inference compute costs manageable. GGUF is a binary file format created by the llama.cpp project that packages model weights, tokenizers, and metadata into a single portable file for efficient local execution.