Micius-Flash: Lightweight Agentic Intelligence for Engineering

Micius-Flash. Lightweight Agentic Intelligence for Engineering. By Hongcheng Gao and twenty-five others. Published on the fourth of November.

The engineering targeted here spans semiconductor knowledge, hardware coding, multimodal understanding of semiconductor material, and agentic work through specialized tools. The title promises lightweight agentic intelligence: one model, not a large one, that covers all of those at once.

Micius-Flash is a model for engineering, with twenty-seven B parameters. It takes multimodal input natively, reasons deeply, and has a context window of two hundred and fifty-six thousand tokens. The strategy for building it is to train several experts first and then fold them into one.

Data comes from the Micius pipeline, designed as a data flywheel driven by a knowledge tree. It runs in three stages: constructing and expanding the knowledge tree, synthesizing data, and validating quality while augmenting with hard examples. From that data the team trains expert models for semiconductor knowledge and reasoning, for hardware coding, and for semiconductor multimodal understanding, plus agentic experts.

Those experts are then merged through Multi-Teacher On-Policy Distillation, or MOPD. Training is supervised on sequences that Micius generates itself, and in this way the capabilities of the separate experts are integrated into a unified model.

The evaluation spans more than twenty datasets with about fifteen thousand records. On hardware code benchmarks, Micius-Flash averages sixty-four point two percent, which is fifteen point zero points above a baseline of the same size. On a semiconductor-specific agent benchmark it scores sixty-two point four percent, ahead of the same-size baseline by seventeen point two points.

For industrial use the authors also build Micius Harness, which connects the model to specialized tools, among them EDA tools and KiCad, and to execution environments. Case studies in PCB design, chip design workflow automation and software development show the model using feedback from tool execution for multi-step planning, code debugging and iterative verification.

The recipe has two halves that meet in the middle. Specialists are grown from a knowledge tree, one per kind of engineering skill, and a distillation step trained on the model's own outputs gathers them back into a single network. The harness then puts that network in front of real engineering tools. Micius is released as an open-source model, intended to advance agents for science and engineering.