life's recent timeline updates

qwen3.8-27b-nvfp4-5090-laptop-24gb-optimal Qwen3.8-27B TWIN-TURBO (NVFP4 GGUF, all-Q8_0 precision chain) tuned to its limits on RTX 5090 Laptop 24GB: 74-78 tok/s with built-in MTP + CUDA graphs, overthinking -93%, vision 3.9s, 160K stable context (cliff mapped). 10 rounds of measured tuning, rejected approaches documented, continuously updated.

qwen3.8-flash-next-strata-5090-laptop-24gb Qwen3.8-Flash-Next 177B MoE on RTX 5090 Laptop (24GB VRAM + 64GB RAM): llama.cpp 22-25 tok/s -> Strata 93.5 tok/s (3.7x), GPU+CPU both saturated. 7 quant tiers screened over 26 rounds, MTP corruption found & fixed, 256K context + vision. 中英双语实测实录

comfyui-5090-laptop-minimax-h3-wan2.2-nvfp4 RTX 5090 Laptop (24GB) + ComfyUI: 12 model lines / 21 measured configs — MiniMax H3, Wan 2.2, Qwen-Image 2512, HunyuanVideo 1.5, FLUX.2 Klein, Z-Image-Turbo, LTX-Video, ACE-Step 1.5, YuE2, Stable Audio 3, Hunyuan3D 2.1, ERNIE-Image. A controlled A/B proves NVFP4 beats INT8. Bilingual 中文/English, licence-verified.

keel3d KeeL 3D — WebGPU browser game framework (Specul). Samples: FPS, tower defense, open world.

Qwen3.8-Flash-Next-DGX-Spark Deploy the quantized version of the 180B hybrid-MoE model on a single NVIDIA DGX Spark (GB10, 128 GB unified memory, no multi-GPU, no cluster) and serve it stably

life V2EX member #6109, joined on 2011-03-02 13:31:44 +08:00