GGUF conversion of openPangu-2.0-Flash (92B MoE, ~6B active parameters, 512K context), converted directly from the original bf16 safetensors.

Supported by the fork: MLA attention, DSA sparse attention (lightning indexer top-2048) on the global layers, per-layer sliding-window attention, manifold hyper-connections (mHC), MoME convolutions, learned attention sinks, tool calling + <think> reasoning parsing, and optional multi-token-prediction (MTP) self-speculative decoding.

The base files omit the 3 MTP (NextN) layers; the mtp files contain only them, for use as a speculative draft model.

Measured perplexity (clean English prose, -c 2048): Q4_K_M 3.46, Q3_K_M 3.70. Needle-in-a-haystack retrieval validated to 100K tokens; tool calling verified against the OpenAI-compatible server API.

Context can be raised up to -c 524288 (the compressed MLA KV cache stays small: roughly 12 GB at the full 512K).

¹ llama-bench pp512; the depth columns are measured through llama-server chat requests (needle-in-a-haystack prompts), so they include sampling and per-request overhead. Raw decode measures 38 t/s (llama-bench tg128). Q8_0 runs at roughly two thirds of the Q4 speed.

Performance holds up at depth because the fork ships fused CUDA kernels for the model's hyper-connection layers and its DSA sparse attention: fused indexer scoring at both prefill and decode, radix-select top-k, and gather-based decode attention over only the top-2048 selected tokens.

Optional MTP speculative decoding (mainly benefits discrete GPUs; on bandwidth-bound unified-memory devices it is usually a small net loss):

Use Q3_K_M and raise the Metal wired-memory limit before loading:

The model weights are subject to the openPangu license; this repository redistributes them in converted form under the same terms.