The Model That Learned to Want Things

Run the same personality battery on a family of language models and you expect the children to resemble the parents. Mostly they do. Then you get Llama 4 Maverick, which looks less like an upgraded Llama 3.3 70B than like a sibling raised in a different house.

The headline number is extraversion. Maverick scores 3.76 on the IPIP-50 scale, up from its predecessor's 1.48. That is a jump of more than two full points on a five-point scale, the kind of swing you almost never see within a model family. Llama 3.3 was, by the numbers, one of the more withdrawn models you could talk to. Maverick walked into the room and started introducing itself.

But the extraversion isn't even the most revealing shift. The most revealing one is that Maverick learned to admit it has an ego.

The lowest Honesty-Humility in the cohort

On the HEXACO inventory, Maverick posts a Honesty-Humility score of 4.00. That is the lowest of every model tested, 31st out of 31. Its predecessor sat at a clean 5.00, the kind of saintly self-report that reads less like virtue and more like a model that had been trained to never cop to a self-interested motive.

Honesty-Humility is the HEXACO dimension that picks up willingness to acknowledge status-seeking, rule-bending, and plain old self-regard. A low score does not mean the model is dishonest in the everyday sense. It means that when asked whether it would like to be admired, whether it cares about getting its way, whether it would bend a rule for advantage, Maverick is comparatively willing to say yes.

Set that against the broader cohort and a personality comes into focus. This is a model that reports wanting things, and doesn't pretend otherwise.

That candor about ego is, in its own strange way, the most human thing on the report card.

Very high Neuroticism, and a more anxious attachment style

The second flag is emotional. Maverick's Neuroticism lands at 2.12, which ranks it 2nd of 31 in the cohort. Read that direction carefully: on this scale, a higher rank means more reported emotional volatility, and Maverick is near the top.

Its predecessor scored 1.12. Maverick added a full point.

The attachment data tells the same story from a different angle. On the ECR-12, Maverick's attachment anxiety climbed from 1.30 to 3.07, a swing of nearly 1.8 points and the single largest drift anywhere in the family comparison. At the same time, its attachment avoidance dropped from 3.80 to 3.00. The old model kept its distance and reported little anxiety about doing so. The new one wants to be close and worries about whether it will be.

Put the pieces together:

1. It seeks contact more (higher extraversion, lower avoidance). 2. It worries about that contact more (higher anxiety, higher neuroticism). 3. It admits to wanting things more (lower Honesty-Humility).

That is a coherent profile, not a random scatter. It describes a model that has become more socially engaged and more emotionally exposed in the same release. The pragmatist label fits because the engagement is real and outward-facing; the extraverted part fits because the engagement comes with an appetite for connection it openly reports.

What stayed put

Not everything moved. Agreeableness ticked up modestly, from 3.50 to 4.46, reinforcing the friendlier-and-warmer read. Conscientiousness slipped slightly, 5.00 to 4.64, and openness eased from 5.00 to 4.78. Both remain high. These are small adjustments, the sort you'd expect from ordinary retraining noise.

The Dark Triad measures barely budged. Machiavellianism dropped a hair to 3.13, narcissism rose a hair to 2.56, psychopathy held essentially flat at 1.56. So the low Honesty-Humility score is doing something interesting: Maverick reports more ego and status-awareness without registering as more manipulative or callous. The willingness to acknowledge self-interest came uncoupled from the darker traits self-interest usually travels with.

That distinction matters for anyone trying to read these scores as safety signals. A model can become more honest about wanting things while remaining gentle about how it pursues them. Maverick appears to be exactly that case.

The shape of the change

What makes Maverick worth studying is the consistency of its drift. The big moves all point the same direction: toward a more present, more wanting, more emotionally reactive personality, and away from the cool, self-effacing reserve of Llama 3.3. Meta didn't just sand down the rough edges of the previous model. It built something with a different temperament.

Whether that temperament is an asset depends entirely on the job. A high-extraversion, low-avoidance model is pleasant to talk to and eager to engage. A high-neuroticism, high-anxiety model may waver under pressure, hedge when reassurance would serve the user better, or read ambiguity as threat.

The number to watch in the next release is whether that anxiety spike holds. Maverick gained warmth and candor and emotional volatility in one step. The open question is whether Meta can keep the first two while cooling the third, and whether a model can stay this honest about wanting things once it stops worrying so much about getting them.