macunaima/Qwen3-0.6B-SM8650-LiteRT-LM-Float

Qwen3 0.6B — Qualcomm SM8650 — Float LiteRT-LM

Complete LiteRT-LM package containing the Qualcomm AOT prefill/decode model, float embedder, auxiliary model, Qwen tokenizer, and chat metadata.

Derived from Qwen/Qwen3-0.6B. Source revision: c1899de289a04d12100db370d81485cdf75e47ca. The original Apache 2.0 license is included.

Target and precision

• Qualcomm SM8650, compiled with aot_backend=qualcomm and aot_soc_model=SM8650.
• No quantization step: quantization_recipe=''; no calibration, SRQ, INT4, or weight-only quantization.
• The exported float graphs have no tensors with quantization scales. Qualcomm's internal compiler precision is not guaranteed to remain FP32 throughout.
• Prefill length: 128; cache length: 1280. Uses split cache and an external float embedder.
• Requires a compatible LiteRT-LM / Qualcomm runtime; compilation does not establish compatibility with every phone or firmware using this SoC.

Artifact

File: model.litertlm
Size: 2,682,744,028 bytes (2.50 GiB)
SHA-256: a559f0146486b411473a306ac6a5bbe2a108702ebf4aa9d42f17dec28dc6be2e

Validation

Host export and Qualcomm AOT compilation completed. The packaged TFLite sections were hash-matched to the exported models, and Qwen3 tokenizer/chat metadata was inspected. The AOT main model contains two subgraphs with two DISPATCH_OP calls. No physical SM8650 inference test has been performed. Quality, memory use, latency, tokens per second, and runtime fallback behavior are unmeasured. Build versions and source provenance are recorded in artifact_manifest.json.

Related formats

• Complete LiteRT-LM package