Qwen 3.8 is a large language model in Alibaba's Qwen family. Recent coverage has focused on local use: one user described an $18K four-GPU AMD Radeon R9700 rig serving Qwen 3.8 at about 900 tok/s decode with 16 concurrent requests; another ran the 27B model at 4-bit with a 100K context on an RTX 3090 for 63 hours; a Saluki 27B post claimed about 96% of its performance at roughly one-seventh the size.
4x AMD R9700 local LLM rig for 10-person startup costs $18KA user shared a local inference build for a 10-person startup costing $18K, with a Threadripper 9970X, 128GB DDR5 ECC, and 4 AMD Radeon R9700 AI Pro GPUs (128GB total VRAM). Serving Qwen 3.8 and other models via a Radiance fork, it reached about 6.3-6.8K tok/s aggregate prefill and 900 tok/s decode at 16 concurrent sessions.
Saluki 27B claims 96% of Qwen 3.8 performance at 1/7 sizeA LocalLLaMA user posted about the Saluki 27B model, claiming it reaches about 96% of Qwen 3.8's performance at roughly one-seventh the size, and asked the community for feedback.
User says uncensored models reduce work headachesA user said Qwen 3.8 and Muse Spark 1.3 often refuse tasks, and switching to a fine-tuned uncensored model makes work smoother. The user reported little performance drop and asked whether others had similar experiences.
Did Alibaba abandon its 35B A3B model?A LocalLLaMA user questions whether Alibaba has abandoned its 35B A3B model, noting no new MoE model came with Qwen 3.8 and no small MoE models were announced at Apsara.
Qwen 3.8 27B ran 63 hours on an RTX 3090A Reddit user ran a 4-bit quantized Qwen 3.8 27B on an RTX 3090 with a 100K context window for 63 hours (over 50M tokens) to attempt the Riemann Hypothesis; it didn’t solve it but produced internal memory and strategy logs published on Hugging Face.
Set Up OpenCode with Ollama and sbx on MacA how-to describing setting up a local LLM development environment on Mac using Ollama, OpenCode, and Docker sandboxes (sbx), recommending Qwen 3.8 and Gemma 4 and providing install/config steps.
The A.I. Daily