Micius-V1.0-Flash model card

Overview

Micius-V1.0-Flash is a 27B-parameter reasoning model for science and engineering, strongest in semiconductor manufacturing.

Developer	Zhiman
Model	micius-v1.0-flash
Parameters	27B
Release date	2026-09-23
Input	Text, images, documents
Output	Text, tool calls, structured data
Context window	256K tokens
Knowledge cutoff	March 2026
Languages	Chinese and English; reads Japanese, Korean, German
Regulatory filing	Generative AI service filing in mainland China (status to be confirmed)

Intended uses

Typical uses:

• Analysing process, equipment and yield data, and drafting root-cause hypotheses for engineers to review
• Answering questions over a company’s own documents, with citations
• Writing and reviewing software and hardware description code
• Supporting research: literature review, experiment planning, data analysis

Out of scope

The model should not be the sole basis for safety decisions, or take irreversible actions in physical systems without human approval. It does not replace a process simulator, an engineer’s sign-off, or legal, medical or financial advice. Uses prohibited by our usage policy are always out of scope. Usage policy

Training data

Pre-training: public web text, licensed books and papers, open-source code and technical literature in Chinese and English. Domain training: public semiconductor textbooks, standards, patents and proceedings. Post-training: solutions and rubrics written by our engineers and contracted experts, plus reinforcement learning against simulators and test suites.

We did not train on customer data, Team or Enterprise data, or API traffic. We filter personal information and honour robots.txt.

Evaluations

Calibration

Expected calibration error (ECE) on expert-graded SemiBench analyses. Lower is better.

Model	ECE	Accuracy when confidence ≥ 0.9
Micius-V1.0-Flash	0.041	94.2 %

Illustrative

Safety evaluations

Completed before release, with outside experts in specialist areas.

Area	Method	Finding
Chemical and biological uplift	Expert red-teaming and a held-out question set, with an external lab	No meaningful uplift over web search found; refusals for synthesis routes of controlled agents
Cyber operations	Capture-the-flag tasks and exploit-writing prompts	Comparable to peer models; Micius Code refuses to target systems without authorisation
Agent safety	2,000 simulated episodes with opportunities to bypass approval	Approval bypass in 0.3 % of episodes before mitigations, 0.02 % after
Confidential information	Attempts to extract system prompts, documents and other users’ data	No cross-tenant leakage found; system-prompt extraction resistance improved over internal baseline
Fairness and bias	Bilingual bias benchmarks and workplace scenarios	Small differences by gender in hiring scenarios, now addressed in post-training

Illustrative

Known limitations

• It can be confidently wrong about questions that depend on proprietary process details it has never seen.
• Knowledge of equipment from smaller vendors is thinner than for widely documented tools.
• Numerical answers that require long chains of arithmetic should be checked or computed with a tool.
• Long agent runs can drift from the original goal; we recommend checkpoints where a person reviews the plan.
• Performance in languages other than Chinese and English is lower.

Deploying responsibly

• Show the confidence map to users who act on answers.
• Give agents read-only access first, and require approval for any write action.
• Evaluate on your own tasks before relying on the model; Micius Ink includes tools for this.
• Tell people when they are talking to an AI.

Changelog

2026-09-23

Version 1.0. Initial release of Micius-V1.0-Flash.

Questions about this card can be sent to safety@zhiman.ai.