Atomic Chat HQ ran an identical prompt suite across four quantization tiers of Qwen 3.8 27B to evaluate voxel island generation.
Their benchmark covered complex environments, including cliffside villages, autumn roads, and steampunk sky islands. The tech community expected
PERPLEXITY 🔥: Perplexity Pro and Max subscribers can now use a new Portable Computer on DGX Spark from NVIDIA, powered purely by local models!
Available models: PPLX 27B is a new post-trained model from Perplexity. Qwen 3.8 27B Nemotron 3.5 Lightning (coming soon) They
Tried running the Qwen 3.8 2-bit model locally on my Book4 Pro running Omarchy✌️
It works, but I got around 1.2 tkn/sec ... way too slow to actually use it (expected).. this is pretty much the limit of what I can run on my current hardware... might just wait for 35B MoE model
Calculate the recent popular models and hardware data: M5 Max Qwen 3.8 27B MLX 4bit (dense model)
Basic short contextPP about 900-925 tok/sTPS 32-33 tok/s ~32k TPS and 28-29 tok/s In most cases, there is a significant improvement after turning on MTP or DFlash. The actual measurement of MTPLX is mostly 55-65 tok/sDFlash2, some people have reached about 70 tok/s M5 Ultra Today…
You have reached the end of the archive
All of qwen38