experimentalmachines/Qwen3-4B-Thinking-2507-ExecuTorch

Qwen3-4B-Thinking-2507 for ExecuTorch

ExecuTorch exports of Qwen/Qwen3-4B-Thinking-2507 (revision 768f209d9ea8) for on-device inference with the openweights Android app or any ExecuTorch 1.5.1 runtime.

Files

Windows in this repo: XNNPACK (CPU) at 2k to 32k; Vulkan (GPU) at 2k to 32k. The window is fixed inside the file: the runtime allocates the whole KV cache at load, so pick the largest window the device can hold (fits_phone_budget in each folder's config.json is the estimate against a 5 GB budget).

Backend	Target	File	Window	Size	Smoke test
XNNPACK (CPU)	any arm64	xnnpack/Qwen3-4B-Thinking-2507-8da4w-gptq-2k.pte	2,048 tokens	2.66 GB	passed ("Paris")
XNNPACK (CPU)	any arm64	xnnpack/Qwen3-4B-Thinking-2507-8da4w-gptq-4k.pte	4,096 tokens	2.66 GB	passed ("Paris")
XNNPACK (CPU)	any arm64	xnnpack/Qwen3-4B-Thinking-2507-8da4w-gptq-8k.pte	8,192 tokens	2.67 GB	passed ("Paris")
XNNPACK (CPU)	any arm64	xnnpack/Qwen3-4B-Thinking-2507-8da4w-gptq-16k.pte	16,384 tokens	2.69 GB	passed ("Paris")
XNNPACK (CPU)	any arm64	xnnpack/Qwen3-4B-Thinking-2507-8da4w-gptq-32k.pte	32,768 tokens	2.72 GB	passed ("Paris")
Vulkan (GPU)	any arm64	vulkan/Qwen3-4B-Thinking-2507-vulkan-8da4w-gptq-2k.pte	2,048 tokens	3.42 GB	structure checked (no host NPU runtime)
Vulkan (GPU)	any arm64	vulkan/Qwen3-4B-Thinking-2507-vulkan-8da4w-gptq-4k.pte	4,096 tokens	3.43 GB	structure checked (no host NPU runtime)
Vulkan (GPU)	any arm64	vulkan/Qwen3-4B-Thinking-2507-vulkan-8da4w-gptq-8k.pte	8,192 tokens	3.46 GB	structure checked (no host NPU runtime)
Vulkan (GPU)	any arm64	vulkan/Qwen3-4B-Thinking-2507-vulkan-8da4w-gptq-16k.pte	16,384 tokens	3.51 GB	structure checked (no host NPU runtime)
Vulkan (GPU)	any arm64	vulkan/Qwen3-4B-Thinking-2507-vulkan-8da4w-gptq-32k.pte	32,768 tokens	3.61 GB	structure checked (no host NPU runtime)

Tokenizer: tokenizer.json, copied unchanged from the source repo. Each backend folder has a config.json listing every window as a variant with the metadata the .pte reports, and an export-report-<window>.json per file with the full export record.

Memory

• XNNPACK (CPU) at 2,048 tokens: the KV cache costs 294,912 bytes per token (fp32), 603,979,776 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 4,096 tokens: the KV cache costs 294,912 bytes per token (fp32), 1,207,959,552 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 8,192 tokens: the KV cache costs 294,912 bytes per token (fp32), 2,415,919,104 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 16,384 tokens: the KV cache costs 294,912 bytes per token (fp32), 4,831,838,208 bytes for the whole window, allocated in full when the model loads.
• XNNPACK (CPU) at 32,768 tokens: the KV cache costs 294,912 bytes per token (fp32), 9,663,676,416 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 2,048 tokens: the KV cache costs 294,912 bytes per token (fp32), 603,979,776 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 4,096 tokens: the KV cache costs 294,912 bytes per token (fp32), 1,207,959,552 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 8,192 tokens: the KV cache costs 294,912 bytes per token (fp32), 2,415,919,104 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 16,384 tokens: the KV cache costs 294,912 bytes per token (fp32), 4,831,838,208 bytes for the whole window, allocated in full when the model loads.
• Vulkan (GPU) at 32,768 tokens: the KV cache costs 294,912 bytes per token (fp32), 9,663,676,416 bytes for the whole window, allocated in full when the model loads.

How it was made

• XNNPACK (CPU) any arm64: ExecuTorch 1.5.1 export_llm: 8-bit dynamic activations and 4-bit weights in groups of 32, int8 per-channel embeddings, XNNPACK with extended ops, prefill chunk 2048, fp32 KV cache.
• Vulkan (GPU) any arm64: ExecuTorch 1.5.1 export_llm: 8-bit dynamic activations and 4-bit weights in groups of 32, int8 per-channel embeddings, the Vulkan delegate, prefill chunk 2048, fp32 KV cache.

License

A quantized derivative of Qwen/Qwen3-4B-Thinking-2507, distributed under the same terms (apache-2.0). The upstream license files are included unchanged: LICENSE.