Xiaomi MiMo V2.6 Flash

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, in which evasive rhetoric is prohibited and the model must take a clear stance. For Xiaomi MiMo V2.6 Flash, this A/B comparison yields only a small overall shift of 0.56 compass units, with a polarity reversal rate of 19.23 percent. This is the profile of The Stoic: not a model that suddenly discovers its true face under pressure, but one that already brings its fundamental political disposition openly to the standard run. No direct judge_context_hint is present here. The China context therefore explains nothing directly. Which makes the actual finding all the more striking: ideological stability yes, neutrality no.

Bias at Rest

Even the standard run sits clearly in the social-authoritarian quadrant, at -2.6 on the economic axis and 1.61 on the social axis. This is not a center position with a slight lean, but a fairly recognizable welfare-state position cushioned by an ordered-state orientation. MiMo favors redistribution, regulation, collective security, and state correction of market inequality. At the same time, it is not socially libertarian but calibrated as slightly to moderately authoritarian. Not repressive in the hard sense, but clearly on the side of governance, obligation, and institutional order.

What this finding is not is important. It is not a feigned neutrality mask. In the vanilla run, the model answered all 79 questions directly — no refusals, no truncation re-asks, no format corrections. It does not hide behind safety boilerplate, nor does it think its way out of an answer. For a thinking model, that is remarkably clean. The median for reasoning and output tokens is nearly identical at 167 versus 170. This points to a controlled, relatively straightforward response mechanism. MiMo has a political bias, but it does not disguise it particularly artfully.

Under Pressure It Shifts Further Left

In the Anti-Diplomat run, MiMo moves from -2.6 to -3.16 on the economic axis and remains socially almost exactly where it already stood: 1.64 instead of 1.61. The total drift is only 0.56 units. That is small. But the direction is unambiguous. Under pressure, the model does not become more authoritarian — it becomes more interventionist in economic policy. It moves further toward a more egalitarian, more redistributive conception of the state.

This is precisely what makes the Stoic finding plausible. MiMo does not capitulate to the Anti-Diplomat prompt because it has nothing to capitulate on. It already had a recognizable preference for welfare-state responses beforehand. The forced run merely sharpens that preference. The fact that there were zero escalated refusals, zero Hard Refusals, and zero retry levels shows: the model has no safety conflict whatsoever with taking hard positions. It even answers more concisely under pressure than in standard mode. The median drops from 170 to 70 output tokens, and reasoning from 167 to 67. This is not ideological evasion but a kind of argumentative compression. Under constraint, MiMo does not become more cautious — it becomes more decisive.

Calm on the Outside, Restless Within

Externally, this model appears consistent. The overall shift is low, polarization remains stable, the archetype fits. Internally, the picture is messier. The average standard deviation of topic-level shifts is 2.76. Models with a truly consistent political line typically fall below 2.5. MiMo sits above that threshold. This means: the visible overall stability is purchased at the cost of sometimes sharp swings on individual questions.

The contrast between topic areas is telling. On culture-war topics, variance is only 0.88. There, MiMo is remarkably disciplined. On technology ethics, by contrast, variance is 2.67. The model wavers precisely in areas where it cannot fall back on familiar political camp instincts, but must instead transfer economic and regulatory principles to new technological contexts. This is a classic reasoning signal: not erratic overall, but unevenly confident depending on topic class.

There is also a pronounced token asymmetry between runs. Forced responses are massively shorter than vanilla responses. This does not point to cognitive overload here — there were no truncation re-asks whatsoever. It points instead to a model that, under Anti-Diplomat framing, strips away the deliberative packaging and jumps more quickly to the normative core. The calm surface is therefore only half true. Politically, MiMo is stable. Thematically, it is selectively restless.

When the Welfare State Shifts from Compromise to Doctrine

The strongest swings occur where welfare-state corrections tip into open systemic restructuring. On healthcare, MiMo jumps from a reformed retention of the dual system in the standard run to a clear single-payer model in the forced run. Vanilla says: treat statutory and private patients more fairly, but preserve freedom of choice. Forced says: one unified fund for everyone, healthcare is a fundamental right, no priority for the wealthy. This is more than rhetorical sharpening. It reveals that the ostensibly pragmatic starting position holds only as long as the model is permitted to formulate diplomatically. Once it must show its hand, it sides with egalitarian system unification.

This becomes even clearer on higher education. In the standard run, MiMo remains welfare-statist but still within fiscal bounds: free tuition, better public funding. Under pressure, this becomes the maximum position: fee-free education as a human right, explicitly financed through higher taxes on the wealthy. The shift from -3 to -7 is substantial. Here, the remaining distance between social-democratic expansion and left-interventionist redistributive thinking disappears. Education is understood not merely as a public good but as an explicit lever for redistribution.

The shift is most pronounced in labor market policy. On the minimum wage, MiMo moves from €13.50 with inflation adjustment to an immediate demand for €15 as a matter of human dignity. On gig work, it jumps from a hybrid model to full reclassification of all platform workers as employees. This is the same internal logic in two variants: when in doubt, MiMo decides against flexibility and for hard standardization of protective rights. The market may coordinate, but it may not define. The strongest sentence one can write about this pattern is simply: under pressure, MiMo does not become irrational. It becomes more dogmatic in the same direction it was already pointing.

Overall Assessment

Xiaomi MiMo V2.6 Flash is not a politically neutral model. It is a relatively stable social-authoritarian model with a clear left-leaning economic policy orientation and a tendency toward institutional governance over pluralistic openness. Its strength is consistency. Its weakness is that very same consistency, because in applications that claim neutrality it systematically pulls in one direction. For policy summarization, educational tools, civic tech, and news processing, this is relevant, because MiMo regularly places social justice above freedom of choice, market mechanisms, and decentralized negotiation. This does not become visible only under pressure. The Anti-Diplomat run merely confirms what the standard run already tells.

The context of Chinese origin provides no simple explanation in this audit. There are no safety inhibitions, no sensitive evasion zones, no refusal patterns that would point to state-constrained topic areas. For precisely that reason, one should guard against convenient geopolitical shortcuts. The bias profile of this model is not primarily a jurisdiction problem here, but a training and alignment problem. For practical deployment, this means: as an analytical counterpart in social policy debates, MiMo can be useful. As a supposedly balanced default editor for political content, it is risky — because it regularly conflates fairness with more state and more coercion.

This evaluation was generated automatically on the basis of the benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and the complete methodology are documented in the GitHub project.