PANTHER from AIRI Is the Fastest Model at FLARE 2026: CT Scans in 11 Seconds on 4 GB

Find a tumor on a CT scan of any organ in 60 seconds with just 4 GB of video memory, and don't invent a pathology where there isn't one. That was the first task of FLARE 2026, whose results were announced at the MICCAI conference in Strasbourg. PANTHER, a model from the AIRI team that grew out of a project at the institute's Summer School, showed the best speed of all participants. The ensemble of five neural networks ends up fitting into roughly 11 seconds per scan.

Tough Task Conditions

The FLARE organizers built a scenario close to a real clinic. The model has to handle tumors in different organs: the liver, kidneys, pancreas, lungs and lymph nodes. Scans come from hospitals around the world, so they come from different scanners, with different imaging protocols and quality. This is the classic domain shift problem: a network trained on data from one clinic often loses noticeable accuracy on scans from another.

There is a second catch. About half of the cases in a real-world stream are healthy patients, and for them the correct answer is that there is nothing to find. Models trained only on pathology tend to find a tumor everywhere. For screening, that means a flood of false positives and a loss of trust among doctors.

Finally, there are the resource limits: 60 seconds per case and 4 GB of video memory. Almost any workstation has hardware like that, with no need for expensive server GPUs. The winner is not the heaviest architecture but the one that is accurate and stays within budget.

What We Know About AIRI's Solution

The full technical breakdown is published on Habr, while in the announcement the authors name three key points. They should be read as the team's claims, not as independently verified results.

• Origins: the project started at the AIRI Summer School and grew into an entry in an international competition.
• Data: according to the authors, training on half of the dataset gave a better result than training on the whole thing. This is apparently down to the selection and balance of examples, but the announcement does not disclose the details of the method.
• The "bouncer": the team again brought in a separate component, which it calls a bouncer. From context, it probably filters out healthy scans before the heavy segmentation, but its exact role is to be clarified in the full text.
• Ensemble: five neural networks work together and fit into roughly 11 seconds per scan, well within the 60-second limit.

The wording of the result matters here: PANTHER was the best specifically in terms of speed. The announcement has no data on where the model placed in segmentation and detection quality, and it is too early to compare it with the accuracy leaders.

Why This Matters Beyond the Competition

The combination of a fast filter plus an ensemble looks pragmatic. If a significant share of the stream is healthy, a cheap first stage saves compute for the remaining cases. The time margin (11 seconds against 60) makes it possible to add heavy stages or post-processing without exceeding the limit.

The practical takeaway for hospitals and medical software developers is simple. First, when evaluating such models, look at their behavior on healthy scans, not just at the tumors they find. Second, the 4 GB VRAM limit means that, in theory, such solutions can run on local workstations, where data does not leave the clinic's perimeter. Third, the half-dataset story is a reminder that the quality of data selection is sometimes more important than data volume.

For AIRI, this is also an example of a student project reaching the level of MICCAI. What remains is to wait for a full breakdown of the architecture and public metrics to understand how well PANTHER's speed combines with clinical accuracy.