runanywhere

/

qwen3_0_6b_HNPU

like 0

Follow

RunAnywhere, Inc. 8

Adds the v79 (Hexagon v79 / SM8750, e.g. Galaxy S25) bundle alongside the existing v81.

• Same 6-file layout as v81/: prefill (fp16) + decode (W8) + tied lm-head (W8) + host embed (fp16) + manifest + tokenizer, driven by the qwen3_generate host-op.
• Manifest carries a runtime-owned ChatML chat block with Qwen3 thinking disabled (empty <think></think> assistant prefix) — this is what fixes the v81 <|im_end|>-loop / empty-output behaviour for short answers.

Plane A (qhx_generate on-device, S25/v79) — validated:

• What is 2 + 2? -> 2 + 2 = 4.
• What is the capital of France? -> The capital of France is Paris.
• coherent greeting + explanation; prompt ~110 tok/s, decode ~34 tok/s.

This bundle is the Halo-brain Qwen3-0.6B, previously device-validated for tool-calling via the SDK.

PR opened unmerged for Plane-B (real SDK) verification before merge.