Fewer Bytes, More Tokens: How Qwen3.8-27B Decodes Six Times Faster on a DGX Spark

The DGX Spark reads memory at 273 GB/s, which caps plain decoding of a dense 27B model at about 14 tokens per second however much compute sits behind it. Measured on one Spark, Qwen3.8-27B decodes at 11.4 tokens per second without speculation and, for the median request, at 67.8 with the Chisel recipe. Every step between the two either reduces the bytes read per decode step or raises the tokens produced per step, and this post measures each one separately.

October 31, 2026 · 20 min · MdJawad

QuIP#: Achieving Near-Lossless 2-Bit LLM Quantization

QUIP# algorithm for quantizing LLM weights without gradient information.

October 16, 2025 · 28 min · MdJawad

Why Can Your Laptop Run LLaMA? A Deep Dive into Quantization

How 4–8x compression and Hessian-guided GPTQ make 70B-scale models practical on modest hardware—what INT8/INT4 really cost, and when accuracy holds.

October 4, 2025 · 24 min · MdJawad