TII releases Falcon Perception, a 0.6B model for text-prompted object segmentation and detection

On April 1, 2026, TII introduced Falcon Perception, a compact multimodal model with 0.6 billion parameters that can find and segment objects in an image from a plain text query, with no fixed list of classes. For the market this is an important signal: visual understanding tasks that long required complex pipelines of separate blocks can now be handled by one shared architecture, under an open license and without a giant model.

Most modern computer vision systems are still built modularly: one encoder extracts visual features, another block mixes them with text, and a separate decoder then predicts boxes, masks or answers. This works but scales poorly: each new type of error is usually fixed with a new module, and the interaction between language and image stays limited. With Falcon Perception, the team at the Technology Innovation Institute in Abu Dhabi bets on early fusion: image and text enter a single token sequence from the very first transformer layer.

Architecturally, the model is a single Transformer with a hybrid attention scheme. Image tokens see each other bidirectionally and build a global visual context, while text and special tokens are decoded causally, relying on the already processed picture. For each object it finds, the model runs through a short chain of steps: it first determines the center coordinates, then the size, and then builds the segmentation mask. This interface lets it handle a variable number of objects, from zero to hundreds in a single image, and keeps mask generation from becoming too expensive computationally.

There is serious data preparation under the hood. Initialization was done by distillation from DINOv3 and SigLIP2, to combine strong local visual features with better language grounding. The model was then trained on a set of 54 million images, 195 million positive text expressions and 488 million hard negative examples. An ensemble of SAM 3, Qwen3-VL-30B and Moondream3 was used for automatic annotation and filtering, and disputed cases were sent for manual review. Separately, TII introduced PBench, a new diagnostic benchmark that breaks results down by difficulty level: from simple objects to OCR prompts, spatial relations and dense scenes with hundreds of instances.

The metrics look convincing. On SA-Co, one of the open segmentation benchmarks, Falcon Perception scored 68.0 Macro-F1 versus 62.

3 for SAM 3. The gain is especially visible where simply recognizing an object is not enough: attributes and subtypes, queries with text inside the frame, and spatial phrasings such as "the car on the left" or "the third window from the left". On PBench the gap on simple objects is small, but it reaches 21.

9 points on spatial tasks, 13.4 on OCR-guided queries and 15.8 on relational tasks.

The weak spot for now is presence calibration: on MCC the model trails SAM 3, 0.64 versus 0.82, meaning that in hard negative scenarios it still more often gets the "object is absent" answer wrong.

The model also has a practical side. Falcon Perception is released under Apache 2.0, is available on Hugging Face and GitHub, and is designed not only for lab experiments but for production use. Inference relies on a stack built on PyTorch FlexAttention and a paged KV cache; according to the team, on an H100 typical latencies are about 100 ms for prefill, about 200 ms for feature upsampling and roughly 50 ms for decoding several instances. TII also showed that the same early-fusion recipe carries over to OCR: the companion Falcon OCR model, with 0.3 billion parameters, scored 80.3 on olmOCR and 88.64 on OmniDocBench.

The main takeaway is not that TII has released yet another compact vision-language model. What matters more is that Falcon Perception demonstrates the viability of a simpler, more unified approach to visual understanding: one architecture, one shared stack and fewer workarounds between language and vision. If the team improves presence calibration and reduces false positives in hard negative scenarios, Falcon has a chance to become a strong foundation for assistants, robotics, visual search and interfaces where an image must be understood from human text rather than from a predefined list of classes.