https://cstech.dev Technical writing on AI/LLM systems, full-stack delivery, and open source tooling. https://cstech.dev/articles/turbo8-distilling-qwen-image-2-1-to-8-steps https://cstech.dev/articles/turbo8-distilling-qwen-image-2-1-to-8-steps Sun, 27 Sep 2026 00:00:00 GMT How I distilled Qwen-Image-2.1 into an 8-step LoRA covering text-to-image, editing and transparent images: a first DMD2 run that drifted, the trajectory-anchored rebuild that fixed it, and the evaluation that told them apart. https://cstech.dev/articles/qwen3-8-flash-next-fp8-on-4x-rtx-pro-6000-blackwell https://cstech.dev/articles/qwen3-8-flash-next-fp8-on-4x-rtx-pro-6000-blackwell Wed, 26 Aug 2026 00:00:00 GMT Measured throughput, KV capacity, and long-context reasoning performance of Qwen3.8-Flash-Next FP8 — a 180B-parameter hybrid-attention MoE with 6B active per token — on four RTX PRO 6000 Blackwell GPUs served by vLLM. https://cstech.dev/articles/ornith-1-5-35b-a3b-fp8-on-2x-rtx-6000-ada https://cstech.dev/articles/ornith-1-5-35b-a3b-fp8-on-2x-rtx-6000-ada Tue, 25 Aug 2026 00:00:00 GMT Measured throughput, capacity limits, and two failed reasoning checks from serving Ornith-1.5-35B-A3B-FP8 on two RTX 6000 Ada GPUs with SGLang. https://cstech.dev/articles/qwen3-8-27b-fp8-on-2x-rtx-6000-ada https://cstech.dev/articles/qwen3-8-27b-fp8-on-2x-rtx-6000-ada Sun, 23 Aug 2026 00:00:00 GMT Measured performance, capacity limits, and operational lessons from serving Qwen3.8-27B FP8 on two RTX 6000 Ada GPUs with SGLang. https://cstech.dev/articles/deepseek-v4-flash-on-4x-rtx-pro-6000-blackwell https://cstech.dev/articles/deepseek-v4-flash-on-4x-rtx-pro-6000-blackwell Sat, 01 Aug 2026 00:00:00 GMT Measured performance, capacity limits, and operational lessons from running DeepSeek-V4-Flash on four RTX PRO 6000 Blackwell GPUs.