Micius-Flash opens a 27B engineering model: 64.2% average on hardware coding, 15 points over a same-size baseline

A 27B engineering model now holds semiconductor knowledge question answering, hardware code generation, schematic and waveform understanding, and tool use in a single model, released as an open-source model. Earlier domain-adapted models were typically trained for a single task, while general-purpose models show clear weaknesses in hardware generation, circuit understanding and realistic design tasks; consolidating complementary capabilities into one model risks interference or degradation of general-purpose performance. Evaluated on more than 20 benchmarks with about 15,000 records: 64.2% average on 12 hardware coding benchmarks, 15.0 points above a same-size general-purpose baseline, and 62.4% on a semiconductor-specific agent benchmark, 17.2 points above the same-size baseline.

The model's capability comes from a data production process organized by knowledge structure rather than from scaling general corpora alone. Semiconductor knowledge is fragmented and insufficiently covered by general training data, so data must be collected, synthesized and validated systematically along a knowledge hierarchy. Knowledge trees start from authoritative standards and normative documents, with agents performing retrieval, concept extraction and recursive expansion, plus cross-branch dependencies forming a knowledge graph; synthetic samples are generated from knowledge nodes or local subgraphs, code samples must pass compilation and simulation, and image samples are discarded if a text-only model can answer them.

Capabilities from four expert models are merged into one student through multi-teacher on-policy distillation. Knowledge and reasoning, coding, multimodal and agentic teachers are fine-tuned independently, then the student receives token-level distributional supervision from the teachers on trajectories it generates itself. Distillation uses a forward KL objective that keeps each teacher's top-32 token probabilities and aggregates the remaining mass into one tail category; each trajectory is supervised only by the teacher for its task category, and trajectories are regenerated online as the student policy evolves.

Evaluation covers more than 20 benchmarks with about 15,000 records across semiconductor knowledge, hardware coding, multimodal understanding, scientific reasoning and agentic tasks; case studies show the model using tool feedback in chip design, industrial software development, simulated experiments, PCB engineering, game development and GPU kernel optimization. The authors state plans to explore more comprehensive industrial scenarios and scale Micius to larger model capacities.

The authors report that relative performance varies across tasks: the model ranks second on all three agentic engineering benchmarks, behind GPT-5.6 Sol, and scores 49.70% on SemiBench, below several compared models. The case studies are representative examples, and the authors give no statistical conclusion across scenarios.