Zyphra — Complete End-to-End Pipeline

Zyphra's portfolio spans the entire AI stack: EEG → LLM (dense, MoE, Mamba) → TTS → voice cloning. 1bit supports all of it — 11 models in 1BP format, plus EEG and TTS pipelines documented for ecosystem completeness. This is the family the engine was tuned against, and the one that powers the JARVIS pipeline.

Legend: 🧠 LLM · 👁️ vision · 🗣️ voice · 🧬 EEG · 🏁 end-to-end validated

Models

¹ ZAYA1-8B 1BP is ~6.6 GB full-weight — the 149 MB entry on HF is MoE-expert-stripped; use the complete file. ² Earlier catalogs listed a 739 MB 1BP for the 74B — that is physically impossible for a full 74B model (even ternary ≈ 18 GB) and refers only to a stripped-expert variant; the runnable 74B Q4_K_M GGUF is 45.8 GB. ³ 74B measured on ROCm TheRock HIP 7.15a, zaya-llama.cpp/build-hip (Juste-Leo2 Zaya branch), full GPU offload — see benchmarks/RESULTS-zaya1-74b-benchmark-2026-08-05.md.

Architectures in this family

• Zaya1 — MoE + CCA. Zyphra MoE architecture with Cross-Channel Attention + MoE FFN. Flagship 1BP ternary model. Tile8 GEMV (28-layer, Zaya1-8B shaped) measured at 77 tok/s on ROCm HIP. TQ2 ternary weights bridged to INT8 for NPU via ternary_npu_bridge.h (pack_tq2_to_npu_int8()) onto existing INT8 xclbin kernels — a true native 2-bit AIE microkernel (4× less DDR traffic) is designed but not yet the default path, see the NPU ternary roadmap. CPU AVX-512 portable path ~2.5 tok/s (8B-shaped).
• ZR1 — dense reasoning. Reasoning-tuned dense transformer (Qwen2 architecture). End-to-end validated ~26 tok/s on Vulkan ZINC; 1BP conversion complete. Builds via the Qwen3-0.6B xclbin stanzas (same tile template).
• BlackMamba — Mamba1 + top-1 MoE. No attention mechanism — alternating SSM scan and MoE FFN per layer. The engine's fastest end-to-end family (79.4 tok/s at 1.5B). SSM scan via kernel/ssm_selective_scan.cc.
• Zamba2 — Mamba2 hybrid. Mamba2 SSM layers with sparse attention every 6 layers. ~30 tok/s at 2.7B on Vulkan ZINC; Mamba2 decode block benchmarked at 1270 tok/s on ROCm HIP. NPU AIE2 selective-scan kernel (per-head d_state=64 vectorized, 32 heads/tile).
• Zamba — Mamba1 hybrid. Original Zamba-7B-v1: Mamba1 SSM + shared attention layers. GGUF through ROCm HIP.

Full pipeline depth

• 🧠 LLM — Zaya (MoE+CCA), ZR1 (dense reasoning), BlackMamba (Mamba1+MoE), Zamba (Mamba1/2 hybrid)
• 👁️ Vision — ZAYA1-VL-8B (built-in vision encoder)
• 🗣️ Voice — voice cloning pipeline (RVQ-VAE codec + QLoRA adapter + ONNX decoder + streaming + persona), see JARVIS pipeline
• 🧬 EEG — ZUNA1.1 / ZUNA (diffusion autoencoder, not 1BP) → Zyphra/ZUNA1.1
• 🗣️ TTS — Zonos-v0.1-hybrid / ZONOS2 (neural audio codec + MoE, not 1BP) → Zyphra/Zonos-v0.1-hybrid

1BP catalog

See also: full model support detail · benchmarks SSOT · all families