Macrostiff Model (1 Jan 2026) [General/Model Discussions] model_drop

[willow pine]

SpysyWeeb

Macrostiff Model (1 Jan 2026) [General/Model Discussions] model_drop

How is this model?
poll

[maiden hornet]

Hey I like this model so far. Cheers 🥂

[minor swift]

woah

new model

[nimble torrent]

The name is great at least!

[maiden hornet]

Stop the presses, this model CAN stop on a red

[gritty rapids]

Would you say it’s a smooth stop, or more rough at the very end of the stop

[maiden hornet]

it's real smooth because it still acts like previous models with the creep at the end, went over the crosswalk then finally came to a stop

[stable lion]

drove with this for about 2 hours, it's pretty good no serious regressions anywhere compared to default

long is great on city streets, stops at reds and sometimes goes on greens (albeit slowly)

lat about the same as before which is not a bad thing

highway long sucks though, can't maintain speed

i'll keep it as my new default i think but switch to non exp on highways

which i was doing anyway so overall an improvement

[stable lion]

lateral still likes to crash into parked cars a lot though

but so does default

really need to get back on mainline SP so my firehose data goes up

[tiny estuary]

@narrow valve Have you tested this model yet? If so, any review compared to WMI V7?

[digital stump]

Does the learned steering lag stay regardless of the model used?

[quick ridge]

It does but it occasionally drifts based on the model. I don’t remember which models it was (I think firehose and Nevada) stabilized at different levels of lagd.

[nova arch]

Is there any reason these new models aren't available in staging?

[quick ridge]

Updates aren’t happening to staging while the raylib rewrite is happening, new modeld is needed I believe

[nova arch]

Yeah I'm aware of the rewrite, but the models are downloaded from the cloud. Modeld update makes sense, but I thought maybe some of the new ones should still work. Guess I'll have to switch to dev then.

[quick ridge]

The sunnylink UI lets you adjust almost everything needed. But you can’t adjust the metrics under the chevron or the advanced data on the display right now so have those set before you update.

You may be able to SSH to make the changes but I haven’t tried that

[narrow valve]

I’ll attach the E2E Long (experimental mode) graphs included in the pull request here for record-keeping.

[tiny estuary]

So far I’m liking this model more than WMI V7 for my 2019 RAV4 ICE

[sharp patio]

Does it cut corners like all the other WMI models. WMI V3 seems to be the best at staying on the road during right turns. I think the snow looks like a lane line and it hits the rumble strips.

[tiny estuary]

I feel like it’s similar but the recentering is faster so it feels more consistently centered than WMI.

I drive in mostly 50 mph and curvy city roads so centering is the main requirement I need.

It also has extremely strong torque steering for 90 degree left and right turns.

[sharp patio]

I have a couple 55mph sweeping 90 degree turns that some models basically drive off the shoulder. Even when I'm holding the center of the lane I can feel the model pulling pretty hard.

[narrow valve]

I also tested the MacroStiff Model (8c06e95e-d7c0-4fd9-ba02-9f0b6848785e) for about 30 minutes during a clear daytime drive, under conditions similar to my WMI v7 test.
In terms of lane centering, MacroStiff performs just as well as WMI v7. Both models consistently hold the center of the lane, and I couldn’t really tell a difference in lateral performance. They are equally comfortable and feel very well tuned in that regard—excellent work on both.
However, I did notice a clear difference in E2E longitudinal behavior (experimental mode). With the MacroStiff model, there were many situations where acceleration felt insufficient, and I often had to manually apply the throttle. In contrast, with WMI v7, this almost never happened.
This difference in longitudinal performance is quite striking. WMI v7 feels much more confident and natural in its acceleration behavior, which makes a big difference in everyday driving.

[somber jackal]

The following distance it’s way better than WMI v7 I like it so far but at high speed over 100km/h it has some kind of rubber band effect, lateral ok ✅

[whole lark]

Not bad so far

[sharp patio]

Lat hugged the right side of the lane. I recalibrated the camera and it was a lot better. Long is always really slow and seemed to stop late at stop signs.

[narrow valve]

master merged

[minor swift]

nice

it is a good model

[narrow valve]

Macrostiff is a good model, but I personally feel that WMI v8 is far superior, so I’m really hoping that the development of the WMI model will be completed and merged into master.