GitHub ↗

Reference · Coverage

Model families

16 documented families, one canonical format. The engine imports GGUF, ONNX, and the native 1BP, detects the architecture, and picks a kernel path. No config, no glue.

The families

Per-family coverage matrices live in docs/model-families/<family>.md. The set: bitnet-bonsai, deepseek, falcon, gemma, gpt-oss, granite, kimi, laguna, llama, mistral, olmo, phi, qwen, smollm, whisper, zyphra.

1BP format

1BP is the single canonical format; importers exist for GGUF, ONNX, Q4NX, and H1B. The quantization path: Q4NX (4-bit, per-group BF16 scales) to dequantized float32, then re-quantized to symmetric INT8 (per-tensor max-abs) once at startup. Activations are quantized on the fly with a fixed 5.0 / 127.0 scale, empirically validated on Strix Halo.

Zyphra, the engine-tuned family

This is the family the engine was tuned against and the one that powers JARVIS by default.

ModelParams1BP sizeBackend(s)Perf
ZAYA1-8B8.8B6.6 GBHIP / NPU64 tok/s HIP
ZAYA1-74B-preview74B45.8 GB (GGUF)HIP16.7 tok/s
ZAYA1-VL-8B8.8Bn/aHIP (vision)n/a
ZR1-1.5B1.5B781 MBZINC / NPU26 tok/s ZINC
BlackMamba-1.5B1.5B970 MBMamba1 HIP79.4 tok/s
BlackMamba-2.8B2.8B1.8 GBMamba1 HIP46.0 tok/s
Zamba2-1.2B-v21.2B1.1 GBHIP / CPU30 tok/s HIP
Zamba2-2.7B-v22.7B2.4 GBHIP / CPUn/a
Zamba2-7B-v27B6.6 GBHIP / CPUn/a
Zamba-7B-v17B4.3 GBMamba1 HIPn/a
Watch out

ZAYA1-8B 1BP is ~6.6 GB full-weight; the 149 MB entry on HF is MoE-expert-stripped. And the 739 MB 1BP for the 74B is physically impossible for a full 74B model; the runnable Q4_K_M GGUF is 45.8 GB.