Frigate+ 2026.0 Base Model Update

[blakeblackshear]

The base model in Frigate+ has now been updated to 2026.0. All model requests going forward will use this version. If you have a model request that has not finished processing, it has automatically been updated to use 2026.0.

For this release, I tested over a dozen variations of training parameters for YOLOv9 to maximize performance. In my testing against images uploaded to Frigate+, the updated base model has fewer false positives and false negatives. Additionally, I have updated the models section of Frigate+ with a dropdown to filter to the most recent base model to make it easier to manage the growing list of options.

I am planning to add YOLOv9 support for Google Coral devices alongside Frigate 0.17. When that happens, you will need to submit a new model request as the current YOLOv9 models will not be automatically compatible.

I have also added some additional candidate labels requested over the past few months: la_poste, lawnmower, and heron.

[bubba925]

Love the continued updates and features. Requested a model two days ago based on some recent false positives and I'm not even slightly upset I'm requesting one again today.

Just wanted you to know I feel I get more than my fair share out of the frigate plus subscription. Hoping I can contribute to the code base in the future when I've learned more and have more time. Until then, take my money.

[Cold-Lemonade]

Is it just my imagination or are the suggested bounding boxes for labels in Frigate+ tighter than ever before?

[ibrewster]

Just out of curiosity, do we know how the accuracy of the YOLOv9 models running on Google Coral Devices (I suspect where they are running doesn't actually matter) compare to the current mobledet models? My past experience has been that YOLO models are significantly better, but I am curious if this is still the case (my experience is at least a couple of years out-of-date) and if it holds true once the YOLO model is made to work on the Coral?

Of course, once v0.17 is out, it sounds like I'll be able to try it out for myself... :)

[NickM-27]

They are quantized so it is not the same as the YOLOv9 that is running on (i)GPUs, I have seen an number of community users reporting it working with varied ranges of success.

[reflectivecode]

I look forward to the base model update every quarter, and I eagerly await the graduation of the chicken and possum labels. Thank you, @blakeblackshear

How many more chicken photos do I need to submit? 😆

[ProsjektX]

Any chance we will see larger models as well soon, at least m size yolov9?

[nsprawling]

Curious about this, don't know if the impact on inference time would be an issue? Or whether the improvement in accuracy is too minor to be worthwhile

[blakeblackshear]

I may be able to add this at the same time I add support for the Coral. Some of the modifications made there may bring a speed boost to the existing models without sacrificing accuracy, and that would make the larger models suitable for realtime detection.

[nsprawling]

Model already requested was upgraded to the new base, and i've got an immediate noticeable improvement in night time detections of vehicles. Excellent as always.

[KyGuy2002]

Excited to try out this new model.

I am curious if the images we submit to frigate+ contribute to the whole dataset or if they are only for our fine tuned models? Do they go through another review before going into the base model or is the human error low enough from user submitted data that it doesn't need to?

Also, can the public access this dataset for our own uses or is it a private frigate+ database?

Thanks

[nsprawling]

Docs > FAQ > Frigate+ > Models
https://docs.frigate.video/plus/

Frigate+ offers models trained on images submitted by Frigate+ users from their security cameras and is specifically designed for the way Frigate NVR analyzes video footage. These models offer higher accuracy with less resources. The images you upload are used to fine tune a base model trained from images uploaded by all Frigate+ users. This fine tuning process results in a model that is optimized for accuracy in your specific conditions.

[NickM-27]

Blake hand selects specific images from the users uploaded images to include in the base model. You can read previous base model updates where he detailed that process

[blakeblackshear]

In addition to what was stated above, I want to clarify that the terms and conditions and privacy policy for Frigate+ are designed to ensure the privacy of the images you submit. They are never shared publicly or with other users and only ever stored in the specific locations outlined in the privacy policy.

[Cold-Lemonade]

@blakeblackshear How many images total go into the Frigate+ base model now? And how many car and person labels does it have?

[blakeblackshear]

Almost 70,000 images. 55,000 car labels and 25,000 person labels.

[WiredLife]

would it be possible to copy the following labels from the original yolov9 to the frigate+ yolov9? they would be very helpful for crime detection
backpack
handbag
suitcase
bottle
knife

[blakeblackshear]

There is no way to "copy" labels from one model to another. I would need these objects labeled in the Frigate+ dataset, so the process would be the same as any other label. Detecting crime may ultimately be better solved with VLMs through review summaries anyway.

[WiredLife]

i know bottle and knife would be difficult
but backpack, handbag and suitcase would be very nice to have
they have likely no false detections with the default yolov9 i've tested

[zhamm]

backpack, bottle, and knife already part of the coco dataset and should work, at least on the base yolov9 model. I don't think the dataset images for bottle and knife are the kind that would be very helpful (when used by a person as a weapon), however. I don't know if Blake is retraining on the entire coco dataset (why would we need the food items, kitchen items or utensils...). When I trained my yolov9 model I used the entire coco dataset and pared down the labels I didn't need, then added guns, fire, and smoke images/labels.

[lawul]

Looking forward to the heron label beging added into the model.
Will the candidate label be mapped to bird until it's added? If yes, the corresponding indicator is missing in the list of labels on frigate+ website.

So far I have labeled herons as birds. Is it possible to implement a filter option to show only pictures of a camera with bird labels? That would speed up going through the pictures to find the herons and relabel them.

Thanks for the great work!

[blakeblackshear]

Yes! I forgot and will add that in.

[ristomatti]

I reckon this subthread is already off topic, but I've wondered for some time what to do with raccoon dogs detected as e.g. dogs (or rodents detected as bears on floor level cameras)?

So far I've decided not to report them as false positives as I definitely want Frigate to detect all animals. I've deliberately enabled detection of somewhat similar looking species to get recordings of other species even if the enabled ones do not even exist in nature in our country (e.g. raccoons ). I have not reported them as positive matches either to not confuse the model. 🙂

[mburruss]

Heron is a nice addition. My yard is frequented by cranes more than herons (except when herons decide to make messes on my patio). Herons and cranes are visually similar. Should I label the cranes as herons? Should I also label them all as birds (duplicate labels) so that they show up as birds if people choose to not recognize herons but do choose birds?

[Roachforlife]

Prepping to start my journey with Frigate plus so this is exciting. Curious if you have an approximate time line on the coral stuff? This is what I run so if not too far off I may just wait and start prepping images in the meantime. Thanks

[blakeblackshear]

Hard to say until I get further in the process, but I am aiming for within the next few months. The Coral will have an upper limit on the number of labels it can support, so I may have to make some significant changes to the training backend.

[DDomnick]

Question: I submitted a fairly extensive (for me) and my first, fine tuned model request on 1/8/2026 and installed it:

Base Model: 2025.3 (Fine-Tuned) | Train Date: 1/8/2026, 4:30:27 PM
Model Type: yolov9s (640x640) | Supported Detectors: onnx, rocm, openvino, rknn, zmq

If I now select the new base model 2026.0 in the drop down for same model type, are my prior submissions integrated into the new model? Or do I need to submit my fine tunes again?

[hawkeye217]

Your own submitted images are always included in any fine tuned model request. So no, you don't need to do anything else if you want to request a new fine tuned model based on 2026.0.

[nsprawling]

Your annotated images for use in fine-tuning are persistent, and will be used for any future fine tuning requests you make. The base models aren't specifically tuned to your submissions. You should request a new fine tuned model, which will take the latest base model, and fine tune it using your annotated submissions

[dlarkinsPonderosa]

I'm curious as to whether those with experience would recommend waiting until the YOLOv9 works with my Coral or to migrate over to the YOLOv9 on my current integrated GPU now.

I have Frigate installed as per the recommended practice (Docker on a bare metal OS install, no virtualization) with nothing else running. I have 12 cameras mostly (all except my doorbells) running detection at very high (full 4K) resolution (mostly looking for people and animals - bear, deer, cats, dogs, racoons - in very large rural areas). My Coral performance is fine but I do get a decent number of false positives (especially at night) even with a Frigate+ model trained with close to 6000 images.

My GPU is an integrated Iris Xe graphic chipset on an Intel Core i7-1360P (12 cores) in an Intel NUC 13 Pro.

Thoughts or experiences?

[dlarkinsPonderosa]

Is that the 640 or 320?

320x320

[dlarkinsPonderosa]

Just to close this out I did a few things that seem to have approved it somewhat.

1. A bios update on my NUC.
2. Changed the power control settings for PL1 (sustained) and PL2 (burst) from their defaults of 40 and 64 to 64 and 86.
3. Processed updates for all packages at the OS level which I hadn't done for a bit.

I'm back at one detector with a decent inference time. The load on it is still high.

I don't know much about how inference time correlates (if at all) to how effectively the detector is working. If the inference time is decent does that generally imply that the queue isn't backing up etc... or is that just the time it takes from when an input is taken from the queue until it is processed?

[NickM-27]

it is the time it takes from the input of the image to the detector to when we get an output of objects i that image, you can generally cehck the health by looking at the skipped fps on the camera tab

What is your overall system CPU usage?

[dlarkinsPonderosa]

My CPU usage is always fairly high ~60%. I attribute that to the high resolution, high fps streams for record and detect.

Is there any acceptable value for skipped frames other than 0?

Right now I am seeing 55 frames, 26 detect, and 0.2 skipped in the overview for all cameras.
Earlier in the day with more motion the skipped was as high as 7 on a few cameras. I would assume that as long as I don't have fast moving objects I should still catch them in other frames?

[NickM-27]

Skipped fps during high activity is not a problem, you want to avoid skipped fps during typical / average levels of activity as a general rule

But yes the tracking can generally still work well with some skipped fps

[Matzebhv]

Yolo for coral and adding heron. Very cool.

[Matzebhv]

I have added heron to some of my cameras. All pictures of these cameras are unverified now. A "dismiss all" button would be fine, i do not want to klick trough 1000s of pictures because i knew there is no heron on them.

Regards
Matthias

[blakeblackshear]

The current counts accurately represent which images have actually been reviewed by a user for each label. Otherwise, when heron is added officially automatic suggestions would be skipped for all of these older images. I think I just need to adjust the UI so it doesn't feel cluttered by older images that haven't been verified for some labels. It's simply not important to ensure all the older images are reviewed for candidate labels that aren't even supported by the model yet, but the UI makes it seem like it is. It also sounds like there is a bug that needs to be fixed which causes you to flip to old images after finishing the new images.

Here is further discussion on this topic: #19053 (comment)

[thanhtantran]

Is that the label fire and smoke already in model?

[blakeblackshear]

The available labels are listed here: https://docs.frigate.video/plus/#available-label-types

What are you trying to use fire/smoke detection for?

[naitkris]

Wow, very impressed with this new Base Model! I requested a new mobiledet edgetpu model yesterday based on 2026.0 Base and it is working great so far - no false positives yet and picked up person and cat multiple times correctly (previously had false positives every few hours). Running Frigate 0.17.0 Beta 2 on a N100 with 2 x Coral USBs.

[jhmeier]

Absolutely agree. Had a lot of false positives especially in the night - no more since creating a model on the new base. Thanks for the awesome work

[beatmag]

Just wanted to ask what model sizes are available for yolov9?
Is it Tiny, Small, M , L?
Frigate+ has T and S?

Which sizes is supported on Coral Edge TPU( in future) and Hailo 8L?

Any one know Hailo 8L inteference time with Yolov9 S and T?

[zhamm]

Can we add some additional label classes for the frigate+ model? Specifically gun, fire, smoke would be very useful for many users I would think. I successfully trained a Yolov9 model to include those labels. The gun label is tricky since frigate wants to box on the person and doesn't always alert on the gun, but I'm sure someone with more experience doing model training than me could figure this out.

[bagobones]

Just wanted to report that I have been running a fine tuned 2026.0 yolov9s model now for almost a week and I haven't had any bird / bear / dog false positives yet.

Which seems to be a significant improvement over the previous base model fine tune.

[dmptrluke]

When do you think YOLOv9 support for Google Coral will be available on Frigate+? My subscription expires soon and it would be good to get one last training in before it does.

[naitkris]

Eager to hear the same as I am in a similar situation but with 2 trainings and 6 months or so of my current Frigate+ subscription left 😄 Going to renew Frigate+ anyways but eager to try YOLOv9 using Frigate+ on my two Corals.

[joselito11]

I also waiting for this great feature. Will support the frigate team in future, because they make really good work.

[blakeblackshear]

My goal is on or before 4/15 with the next base model.

[ZhaiSoul]

Currently, I am using the Plus model to detect faces, but in face recognition, there are often objects that are not actually faces. Is there a way to improve this?

[blakeblackshear]

That seems very unusual. I think there are some situations where these images get pulled from the wrong frame if the frames are expired from the cache before they can be created. The frame cache is a rotating buffer, so its possible for the underlying data to get replaced when its trying to grab a thumbnail.

[bagobones]

Anyone have garbage_truck actually label instead of car?

I have double checked my outdoor cameras and they all have the garbage_truck label enabled in addition to car but the trucks always detect as a car.

[dlarkinsPonderosa]

I'm not sure the candidate labels work the way you think they do. I think they are available for annotating in frigate+ but they are not actually in the models yet and therefore won't be detected and labeled automatically.

For now, if you label something as a garbage truck it will be automatically trained as a car when you request a new model.

Eventually when garbage trucks are added to the model, then the ones you have already verified would be trained as a garbage truck.

For now, I think the benefit of candidate labels is to help build out a set of images for inclusion in the base model eventually. Of course you also get the benefit of having these in your trained models without having to go back and verify them when the labels are added to the model.

[NickM-27]

Yes, that's correct

[bagobones]

Never mind I thought garbage truck was already out of Candidate.. My mistake,