Falcon — TII Dense (Parallel Attention)
TII's Falcon3 — parallel attention+FFN architecture with multi-query attention (MQA). Broad size range, GGUF on all GPU backends and CPU.
Models
| Model | Params | 1BP Size | Backend(s) | Perf |
|---|---|---|---|---|
| Falcon3-1B | 1B | 675 MB | GGML-Vulkan / ZINC / NPU / HIP | — |
| Falcon3-3B | 3B | 1.4 GB | GGML-Vulkan / ZINC / NPU / HIP | — |
| Falcon3-7B | 7B | 4.0 GB | GGML-Vulkan / ZINC / NPU / HIP | — |
| Falcon3-10B | 10B | 5.7 GB | GGML-Vulkan / ZINC / NPU / HIP | — |
Notes
- NPU: Falcon-7B build stanza (
build_falcon_7b). Uses padded dimensions (H=4544→4608, nearest multiple of 128). Build with./build_xclbins.sh falcon_7b. - GPU HIP: GGUF validated. Vulkan / CPU: functional, perf pending.
See also: full model support detail · benchmarks SSOT · all families