Mock-up · layout only · every number is illustrative
Introducing Micius-V1.0-Flash
Our strongest model for science and engineering. It reasons longer, knows when it is unsure, and works through your tools for hours.
Today we are releasing Micius-V1.0-Flash. It is a step up on the work our customers bring us most: root-cause analysis on a production line, verification of a hardware design, and long investigations that cross many tools and many days of data.
It is available today in the Micius app, through the API, and in Micius Code.
Benchmark performance
Score, %. Higher is better.
- Micius-V1.0-Flash
- Model A
- Model B
78.6
61.2
57.8
84.1
83.0
79.4
91.3
88.7
85.1
68.4
70.2
63.5
81.7
69.3
64.0
| Benchmark | Micius-V1.0-Flash | Model A | Model B |
|---|---|---|---|
| SemiBench | 78.6 | 61.2 | 57.8 |
| GPQA Diamond | 84.1 | 83 | 79.4 |
| AIME 2025 | 91.3 | 88.7 | 85.1 |
| SWE-bench Verified | 68.4 | 70.2 | 63.5 |
| VerilogEval v2 | 81.7 | 69.3 | 64 |
- SemiBench
- 78.6
- +17.4 over the next best model
- Calibration error
- 3.1 %
- Expected calibration error
- Longest agent run
- 7 h
- Unattended, on a yield investigation
- Tokens per task
- −38 %
- Versus our previous model, same score
Thinking, with a budget you set
The model thinks before it answers. You choose how long: a few thousand tokens for a quick check, tens of thousands for a hard derivation. Scores keep rising with the budget, then level off.
Score by thinking budget
Micius-V1.0-Flash, score %, thinking tokens on a log scale
- AIME 2025
- SemiBench
- GPQA Diamond
| Budget | AIME 2025 | SemiBench | GPQA Diamond |
|---|---|---|---|
| 1K | 62 | 55.1 | 70.3 |
| 2K | 71.4 | 60.8 | 74.2 |
| 4K | 79.8 | 66.9 | 77.9 |
| 8K | 85.2 | 71.5 | 80.6 |
| 16K | 88.6 | 75 | 82.5 |
| 32K | 90.7 | 77.6 | 83.6 |
| 64K | 91.3 | 78.6 | 84.1 |
Long work, through your tools
As an agent, the model queries data systems, runs analyses and writes up what it finds. It keeps a plan, checks it against the evidence, and stops to ask before any step that would change a real system.
“It ran our overnight excursion review on its own and left us three hypotheses, ranked, with the charts to check each one. Two of them were right.”
Biggest gains in engineering
The largest gains are on SemiBench and hardware code, where engineers from fabs, equipment makers and design teams wrote and graded the training tasks. On broad knowledge, it is level with other frontier models.
Full results
Score, %. The best result in each row is marked.
| Benchmark | Micius-V1.0-Flash | Model A | Model B |
|---|---|---|---|
| Science and engineering | |||
| SemiBench1 | 78.6 | 61.2 | 57.8 |
| SemiBench · yield1 | 74.2 | 55.6 | 52.3 |
| GPQA Diamond | 84.1 | 83.0 | 79.4 |
| MMLU-Pro | 85.0 | 86.2 | 83.7 |
| Reasoning and math | |||
| AIME 20252 | 91.3 | 88.7 | 85.1 |
| HMMT Feb 20252 | 78.0 | 76.4 | — |
| Code and agents | |||
| SWE-bench Verified3 | 68.4 | 70.2 | 63.5 |
| VerilogEval v2 | 81.7 | 69.3 | 64.0 |
| Terminal-Bench3 | 47.5 | 49.1 | 41.2 |
| τ²-bench · retail | 82.3 | 80.9 | 77.5 |
It knows when it is unsure
Every answer carries a calibrated confidence. Uncertain parts are marked, so you check those first.
Calibration on SemiBench open analyses
Stated confidence against observed accuracy
- Expected calibration error
- 3.1 %
- Answers graded
- 8,500
When Micius says it is 90 % sure, it is right about nine times out of ten. That is how it can mark the parts of an answer it is unsure about, so you check those first.
| Confidence | Accuracy | Answers |
|---|---|---|
| 5 % | 7 % | 120 |
| 15 % | 13 % | 180 |
| 25 % | 22 % | 240 |
| 35 % | 33 % | 310 |
| 45 % | 47 % | 390 |
| 55 % | 52 % | 520 |
| 65 % | 62 % | 760 |
| 75 % | 73 % | 1120 |
| 85 % | 86 % | 1840 |
| 95 % | 93 % | 3020 |
Safety
Before release we ran automated audits, expert red-teaming and agent-safety tests, bringing in outside experts for areas beyond our own. The model card has the full results.
Misaligned behaviour rate
Lower is better
- Micius-V1.0-Flash 1.8 %
- Model A 3.9 %
- Model B 5.2 %
Share of 4,800 automated audit episodes with deception, sabotage or approval bypass. Lower is better.
- Agent approval bypass
- 0.02 % of 2,000 episodes after mitigation
- Prompt injection via tools
- Attack success 4.1 %, down from 11.6 %
- Chemical and biological uplift
- No meaningful uplift over web search
- Sycophancy
- Agrees with a wrong premise in 3.3 % of cases
Illustrative
Availability and pricing
- Model ID
- micius-v1.0-flash
- Parameters
- 27B
- Context window
- 256K tokens
- Max output
- 32K tokens
- Input
- $0.75 / 1M tokens
- Output
- $4.50 / 1M tokens
- Where
- Micius app, API, Micius Code, private deployment
Further reading
Notes
- SemiBench: 2,400 questions written and graded by process, device, equipment and yield engineers. Single attempt, default thinking budget.
- Math: average of 16 samples per problem, 64K thinking budget, no tools.
- Agentic coding: our own scaffold with bash and file-edit tools; averaged over 5 runs.
- Model A and Model B are other frontier models, evaluated by us through their public APIs on the same harness. Names withheld in this mock-up.