Qwen3.8
Chat, draft and summarize with a native desktop language model.
No runnable checkpoint is available
This family has no verified executable variant. Download and inference are unavailable.
- 27B: Pinned llama.cpp b12359 ran the Q4_K_M model and Q8_0 vision projector on this Mac. CPU and Metal passed arithmetic and translation. Both failed default OCR; the 768-image-token Metal retry passed the unchanged OCR fixture. The adapter, full quality/speed matrix, fallback, cleanup and public DesktopAI checks remain pending.
- Flash-Next: Official weights exceed local capacity. MLX-VLM 0.7.12 contains qwen4_exp. The 67.67 GB oQ2 MLX conversion or 67.33 GB IQ1_S GGUF plus projector leaves too little local disk headroom alongside the current 27B artifacts for model and runtime staging. Quantized text/image quality and working memory remain unmeasured.
- 2.4T-A95B: The official FP8 files alone total 2.50 TB. The inspected 1-bit GGUF is about 508 GB, exceeding available local disk and 96 GiB RAM. No alternative execution route was verified on this Mac.