A new open-source model classifies text, code, and file types straight from raw bytes, with no tokenizer in sight.
Researchers released pico-type, a byte-level content classifier built from roughly 1.5 million parameters that reads raw UTF-8 bytes directly, skipping the tokenizer and subword vocabulary most language models rely on. In a single forward pass it predicts seven properties at once: coarse content type, modality, subtype, one of 62 code languages, one of 30 text languages, file MIME type, and whether the content contains sensitive material like API keys, passwords, JWTs, or SSH keys. The architecture stacks convolutional layers with bidirectional attention, then splits into seven separate classification heads sharing one trunk. It ships in four sizes, from a 16-dimension tiny variant to a 576-dimension pro version, with even the largest exporting to an ONNX file under 210 KB and running inference on CPU in under 10 milliseconds.
That speed and size matter because content sniffing (detecting file types, languages, and leaked secrets) usually runs on regex rules and heuristics bolted into CI pipelines and editors. A model this small could slot into that same spot without adding real latency. But the numbers show real-world use is uneven: code language identification hits just 60.3 percent accuracy on an external benchmark, versus 98.2 percent for text language, though both are large jumps of plus 57 and plus 79 percentage points over a version trained only on synthetic data.
A model that nails file-type and language detection but only gets code language right six times out of ten isn't ready to replace your linter's language guesser: it's a fast triage step, not a verdict.