We introduce Micius-Flash, a 27-billion-parameter model for engineering with native multimodal input, deep reasoning and a 256K-token context window. We design the Micius data pipeline as a knowledge-tree-driven data flywheel comprising three stages: knowledge tree construction and expansion, data synthesis, and quality validation with hard-example augmentation. From this data, we train expert models for semiconductor knowledge and reasoning, hardware coding, and semiconductor multimodal understanding, along with agentic experts. Subsequently, we employ Multi-Teacher On-Policy Distillation (MOPD) to supervise training on sequences generated by Micius itself, thereby integrating the capabilities of these expert models into a unified model. On more than 20 datasets with about 15K records, Micius-Flash averages 64.2% on hardware code benchmarks, 15.0 points above a same-size baseline. It scores 62.4% on a semiconductor-specific agent benchmark, exceeding the same-size baseline by 17.2 points. We further develop Micius Harness for industrial applications, connecting the model to specialized tools, including EDA tools and KiCad, as well as execution environments. Case studies in PCB design, chip design workflow automation, and software development demonstrate the model's ability to use tool execution feedback for multi-step planning, code debugging, and iterative verification. We release Micius as an open-source model to advance the development of agents for science and engineering.
Figures & tables
Figure 2: Knowledge-tree-driven data flywheel for semiconductor training data. We build and extend a domain knowledge tree, then map source corpora onto its nodes. Evaluation results point to the nodes where the model is weak, and these guide the next round of data synthesis and refinement.
Figure 3: Overview of the agentic environment and task synthesis pipeline, from real-world scenarios to environment construction, task generation, and trajectory verification.
Figure 4: Overview of the Micius-Flash post-training pipeline. Starting from the base model, SFT produces four specialized teachers for semiconductor knowledge question answering and reasoning, coding, multimodal question answering, and agentic tasks, while initializing the student through supervised warmup. During MOPD, the student generates on-policy trajectories and receives token-level supervision from the teacher assigned to each task category, iteratively consolidating their capabilities into a unified model.
Benchmark
Qwen 3.8-27B
Muse Glimmer
MiniMax M3
Tencent Hy3
Kimi K2.6
DeepSeek V4-Pro
Gemini 3.0 Flash
Claude Sonnet 5.0
GPT 5.6 Sol
Micius Flash
Reasoning and knowledge
SemiBench
42.51
29.94
56.89
52.10
45.51
67.66
68.50
66.40
88.02
49.70
SuperGPQA (Circuits)
70.37
63.89
61.11
73.15
67.59
73.15
72.55
67.86
75.93
72.22
SuperGPQA (Emfield)
84.09
82.95
78.41
86.36
88.64
85.23
87.50
86.67
87.50
86.36
SuperGPQA (Microelec)
76.00
70.00
82.00
78.00
82.00
82.00
78.12
79.31
78.00
84.00
SuperGPQA (Physics)
63.93
63.93
63.93
70.49
72.13
70.49
85.37
72.22
65.57
70.49
Table 1: Results across 25 benchmarks for nine selected reference models and Micius-Flash.
Figure 5: Micius Harness overview. The Capability Broker selects and enforces a scoped tool set; the Engineering Viewer provides artifact-bound inspection; and the Remote Execution Runtime provides local and remote execution, querying saved invocation identities to recover ambiguous calls while leaving unresolved outcomes unknown.
Figure 6: Representative scientific and engineering applications of Micius-Flash. (a) Automated design, verification, and physical implementation of a streaming Sobel edge detector. (b) Development of CAD, 3D modeling, and EDA plugins, together with reconstruction of a 3D modeling application. (c) Autonomous battery characterization through instrument operation and parameter estimation in a virtual laboratory. (d) 3D rendering of an automatically designed soft-latch PCB. (e) Playable Godot games with implemented interfaces and gameplay mechanics. (f) Visual feedback for diagnosing matrix-packing errors during GPU kernel optimization.
We ask whether AI agents powered by locally deployed large language models can reliably automate expert-defined hardware design workflows in an industry-realistic tool-calling setting. In these environments, engineers issue repetitive, dependency-ordered operations---such as creating components, adding ports, and wiring connections---through specialised tools. Confidentiality constraints on component specifications and naming conventions often preclude hosted proprietary APIs, motivating the use of locally deployed models. To study this setting, we build a Model Context Protocol (MCP) server that reproduces the state and dependency logic of a proprietary hardware design tool used in embedded system development and construct a benchmark covering single-operation edits, multi-step dependency chains, invalid requests, misspelled prompts, and multi-server tool contexts. We evaluate seven open-source models comparing pipeline choices including system prompts, tool-description detail, context scope, and single-agent versus multi-agent architectures. Results show that strong models can achieve near-complete expected-call coverage on the benchmarked workflows, but reliability depends strongly on both task structure and agent configuration. Comprehensive tool descriptions consistently reduce failures, few-shot prompting can cause severe inaction for some models, cumulative context harms constrained models, and multi-agent decomposition helps weak workers or long sessions at the cost of additional calls. These findings provide practical guidance for deploying local LLM agents in stateful hardware design environments.
AI agents are increasingly used for simulation-driven engineering. Physical system modeling presents different requirements from general-purpose code generation in software engineering, because correctness depends not only on syntax and executability but also on physical consistency and scenario-dependent behavior. We study this challenge in Modelica, an equation-based modeling language in which a model may compile and simulate while still violating its intended physics or engineering requirements. Across successive revisions, an agent may lose track of requirements or rely on simulation evidence produced by an outdated candidate. To address this challenge, we present Pufibara, an agent harness that maintains persistent engineering state across revisions, associates execution and simulation evidence with the candidate that produced it, and makes submission an explicit agent action. To evaluate end-to-end Modelica agent workflows, we also propose a source-grounded method for constructing realistic and independently evaluable tasks. We use this method to build the 232-task Modelica Agent Workflow Benchmark, spanning Model Repair, Model Generation, and Model Tuning. Each submitted candidate is scored by a benchmark-owned evaluator outside the agent loop. We compare Pufibara with Claude Code as complete harnesses under two matched large language model (LLM) backends. With DeepSeek v4 Flash, Pufibara passes 202 tasks, compared with 185 for Claude Code. With Claude Sonnet 5, Pufibara passes 202 tasks, compared with 187 for Claude Code. Under the repository-reported token accounting, Pufibara records 76.4%-82.5% lower logical-token totals. Its sequential runtime is 6.1%-58.4% lower. These findings show that, even under matched LLM backends, complete agent harnesses can differ substantially in both task success and resource use for physical system modeling.
Embedded board-level development demands extensive expertise, from component selection and device characteristics to interface configuration, driver writing, and system composition. Recently, LLM-based agents have shown strong capabilities, raising the prospect of automating board-level design tasks. However, existing agents either target chip-level design or treat datasheets as static text, offering little guidance on whether a specific component should---or even could---be used for a given board-level design. Moreover, without touching real hardware, they hallucinate incorrect details. To fill this gap, we present {\setlength{\fboxsep}{1pt}blue{white}{DS2Agent}}, an open-source framework that constructs interactive agents from device datasheets. Each agent comprises three models: a knowledge model supporting retrieval and question answering; a digital-twin model capturing the physical functionality of the device; and an interface model supporting driver development and system integration. To suppress hallucinations, we also allow optional connections to real devices, so that the agent can calibrate its internalized knowledge against the device. We evaluate DS2Agent on ten devices spanning seven functional categories. Compared with the baseline, the DS2Agent-augmented agent improves the question-answering score from 87 to 97 out of 100 while using 44.8% fewer tokens. For driver generation, it reduces average token usage per device by 72.9%. For system integration, it reduces total token usage across three FreeRTOS applications by 29.7%.
Hongbin Yang, Chi Zhang, Yinhao Xiao +3
Shandong University · GDUFE-UCM Joint Institute, Guangdong University of Finance & Economic