software-mansion/react-native-executorch-qwen-3.5

Qwen3.5 0.8B and 2B

Instruction-tuned Qwen3.5 text decoders exported for React Native ExecuTorch. Sources: Qwen3.5-0.8B and Qwen3.5-2B. Source revisions and artifact checksums are recorded in the export manifest.

Variants (v0.11.0)

Model	Backend	Weights	File	Size
0.8B	XNNPACK	8da4w	0_8b/xnnpack/qwen_3_5_0_8b_xnnpack_8da4w.pte	525.1 MiB
0.8B	MLX	int4	0_8b/mlx/qwen_3_5_0_8b_mlx_4w.pte	544.0 MiB
2B	XNNPACK	8da4w	2b/xnnpack/qwen_3_5_2b_xnnpack_8da4w.pte	1233.9 MiB
2B	MLX	int4	2b/mlx/qwen_3_5_2b_mlx_4w.pte	1286.5 MiB

All variants use a 2,048-token context and a maximum prefill chunk of one token. Each size uses its own <size>/tokenizer.json and <size>/tokenizer_config.json. The chat template defaults to non-thinking generation and preserves historical prefixes when messages are appended.

Both sizes and backends were tested on an Apple M1 Pro, including logits, generation, and cache rewinds. Recurrent-state restoration is handled inside the exported model.

Compatibility

These files target ExecuTorch v1.4.1 and React Native ExecuTorch v0.11.0. MLX requires physical Apple hardware. XNNPACK runs on CPU.

Use the library's models.llm.QWEN3_5_0_8B or models.llm.QWEN3_5_2B entries. See the downloading models documentation and ExecuTorch's runtime compatibility note.