Micius Team
7 min
Abstract
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.
Sam: That last check prevents an image benchmark from quietly becoming a reading exercise. But how do they turn these different training streams into one model?
Alex: They first train specialist teachers for knowledge and reasoning, hardware coding, visual understanding, and tool-using tasks. The student also gets supervised preparation. Then they use on-policy distillation, meaning teachers supervise responses produced by the student as it currently behaves.
Sam: Why have the student generate the responses? Why not just collect polished answers from each teacher?
Alex: Here, the student generates a response or a sequence of interactions. That sequence goes to the teacher responsible for its task category, which supplies guidance about the next tokens. As the student changes, new sequences are generated, keeping supervision tied to situations the student actually reaches.
Sam: So the report combines an organized data loop with specialist supervision on the student's own attempts. Does it establish which ingredient contributes most?
Alex: Not from the evidence presented in this excerpt. We get the combined system's results, but not a controlled breakdown isolating those ingredients. I'd treat this as evidence for the overall recipe, rather than proof that any particular component explains the gains.
Sam: Let's look at those gains. How broad is the evaluation, and what's the comparison that matters?
Alex: The report evaluates twenty-five benchmarks, covering knowledge, coding, vision, tool-using tasks, and general abilities. Altogether, they contain about fifteen thousand instances. The key comparisons are with Qwen, its same-size general-purpose baseline, and with stronger reference models.
Sam: A specialized model should beat its starting point in its specialty. How large is that improvement here?
Alex: Across the hardware coding benchmarks, Micius averages about sixty-four percent. That's fifteen percentage points above the same-size baseline. The tasks span hardware description languages and engineering programming, rather than just one kind of code generation.
Sam: And when the model has to act through tools, rather than produce a code answer?
Alex: On the team's Semiconductor Agent benchmark, Micius scores about sixty-two percent. That's roughly seventeen percentage points above the same-size baseline. It also ranks second among models with reported results on each of the agent benchmarks evaluated.
Sam: Second place could still leave a substantial gap. How far ahead is the strongest model on that semiconductor benchmark?
Alex: GPT-5.6 Sol scores about eighty-seven percent there. So this is substantial progress over the same-size baseline, not evidence that the smaller specialist generally matches the strongest reference. The report also finds competitive general reasoning scores, though that's different from leading across the board.
Sam: The benchmark results establish a useful improvement. What does an actual repair loop look like when an engineering check fails?
Alex: Take the streaming Sobel edge detector, a circuit that finds edges in incoming images. The model develops the hardware and a separate reference implementation for checking outputs. It then runs electronic design automation tools, the software used to implement and check chip designs.
Sam: Where does that workflow force it to go beyond generating a plausible first design?
Alex: After physical implementation, timing analysis finds paths that take too long. The model reads the reports and revises the hardware, splitting arithmetic across clocked stages and adjusting the control timing. It repeats tests and implementation checks until the reported design meets its timing target.
Sam: That's a concrete example of feedback crossing levels: a physical implementation problem sends it back to the original hardware description.
Alex: And that is stronger evidence of an iterative workflow than a code listing alone. But it's still a case study, not an estimate of success across arbitrary chip projects. The report describes verified design artifacts, not a manufactured chip tested in use.
Sam: Running engineering software also brings operational risks. How does the surrounding system control what the model can do?
Alex: Their Micius Harness restricts tool access according to the project and workflow stage, then enforces that scope when tools run. Its viewers tie supported design files to recorded identities and check their content hashes. For remote jobs, selected uploads require approval, and credentials stay outside the model's process.
Sam: Those are useful controls, but viewing a design isn't the same as verifying it. Does the system preserve that distinction?
Alex: The report explicitly says opening or rendering an artifact does not run engineering checks. It also leaves an uncertain remote outcome unknown if recovery cannot resolve it, rather than blindly submitting the job again. Those mechanisms support traceability without making every result trustworthy by default.
Sam: Given the benchmark gains and the limits of the demonstrations, who should read the full report, and where should they begin?
Alex: Researchers building engineering agents should start with Evaluation. Compare the gains over the same-size baseline with the remaining gap to the strongest model. Then read Data Curation and Post-training for the training recipe, and the Harness description for execution controls. Everyone else can keep this: a useful engineering assistant needs both specialized knowledge and a way to check its own work.
Sam: Keep the checks in view.