One engine, Any model.

The engine reads the model header, picks the architecture and kernel path, and runs. No config, no per-model glue, no registry to maintain. And every family serves from one shared memory pool — NPU, GPU and host alias the same SharedBO pages, zero copies — with one chained API call per layer (the Windows NPU model, matched on Linux).

569architecture tokens

2,030HF arch strings

100%checkpoints mapped

The validated families.

supported & validated functional · perf pending / not yet

Zyphra

flagship · MoE · SSM · dense

1.2B-74B npu · hip · vulkan · cpu Zaya1 · ZR1 · BlackMamba · Zamba/Zamba2 · 6.6 / 45.8 GB 1BP

Qwen

dense · MoE · vision

0.5B-72B npu · hip · vulkan · cpu

Llama

dense

1B-8B npu · hip · vulkan · cpu

Mistral / Pixtral

dense · vision

7B-12B npu · hip · vulkan · cpu

Gemma

dense · embedding

0.3B-4B npu · hip · vulkan · cpu

Phi

dense

3.8B-4B npu · hip · vulkan · cpu NPU + CPU at present · GPU pending

DeepSeek

MoE (MLA)

8B-671B npu · hip · vulkan · cpu

Falcon

dense

1B-40B npu · hip · vulkan · cpu

GPT-OSS

MoE

20B npu · hip · vulkan · cpu

Moonshot Kimi

gated MLA MoE

16B (3B active) npu · hip · vulkan · cpu

BitNet / Bonsai

ternary-native

1.7B-4B npu · hip · vulkan · cpu

Whisper

speech-to-text

V3 Turbo npu · hip · vulkan · cpu Vulkan port pending

Qwen-VL

vision

3B-4B npu · hip · vulkan · cpu ViT not yet on Vulkan

Also validated: OLMo, ZR1, Laguna, Nanbeige, Embedding-Gemma. Per-family pages live in the docs.

1BP format. Dense models use Q4NX 4-bit; only ternary-native checkpoints (BitNet / Bonsai) use TQ2 2-bit. The engine never converts dense to TQ2, because it is quality-destructive and it refuses to do it silently.