OPEN WEIGHTS MODEL
Galvanize-60M
60M parameter prompt injection classifier for agent tool loops. Certified p50 11.52 ms CPU, 1.00 percent tool false positives, 91.60 percent out of distribution injection recall at 8,192 tokens.
11.52 ms
CPU p50 latency
11.52 ms base, 18.18 ms INT8 quantized
1.00%
Tool false positives
1.00 percent (0.67 percent at tau 0.80)
91.60%
OOD injection recall
91.60 percent calibrated on the Deepset suite
8,192
Long context
77 to 97 percent needle recall across 8k
Inside the model
4 layers, 60M params
Distilled stack sized for the agent hot path. Apache 2.0 release.
Native 8,192 token RoPE
No sliding window and no truncation for typical agent prompts with schemas and retrieved docs.
MultiHeadSecurityPooling
4 learned queries over the sequence into 3,072 dims. Heads for overrides, jailbreak framing, delimiter escapes and exfiltration.
Dual distribution
PyTorch model.safetensors 240 MB plus dynamic INT8 ONNX onnx/model_quantized.onnx 176 MB.
Pairs with the deterministic rules layer under 0.1 ms. Open weights cover the semantic engine.
Use Galvanize today
Pull the weights or call the hosted gateway.