Posts

In September 2026 there are 3 models on Hugging Face you would want to put on a pair of DGX Sparks behind Claude Code: DeepSeek-V4.1-Flash at 552B, GLM-5.3-Flash at 320B and Qwen3.8-Flash-Next at 125B. The first leads every benchmark but does not fit in 243 GiB even at 4 bits. The second takes 198 GB and 15 minutes to load. The third fits on a single Spark — apart from the 2 GB it is short. Below: why on this cluster the choice is made by memory arithmetic rather than a benchmark table, why 6 billion active parameters do not give you 90 tok/s, and why for Claude Code what decides it is not tok/s but a single vLLM flag that the fastest config turns off. ...

I own two ASUS Ascent GX10s. Before buying the first one — and again before buying the second — I went through every GB10 machine on the market. This post is the comparison I wish had existed: not a spec-sheet rehash (the spec sheets are nearly identical by design), but a look at where the boxes actually differ, with numbers from months of running one 24/7. The GB10 landscape: one superchip, eight chassis When NVIDIA launched the GB10 Grace Blackwell superchip, it didn’t just build its own DGX Spark — it handed the same platform to OEMs. Today you can buy essentially the same computer from eight vendors: NVIDIA (DGX Spark Founders Edition), ASUS (Ascent GX10), Dell (Pro Max with GB10), HP (ZGX Nano G1n), Lenovo (ThinkStation PGX), MSI (EdgeXpert MS-C931), Gigabyte (AI TOP Atom), and Acer (Veriton GN100). ...