Field-tested ways to run it.
Reproduced on real hardware, with the failure modes worth knowing first.
Production recipes
3Serving DeepSeek-V4-Flash-Vision on 4× DGX Spark: images, DSpark and DCP in one stack
DeepSeek-V4-Flash-Vision-Exp on four GB10 nodes, ported by hand onto the sparkinfer B12X stack before upstream support landed. The interesting part was not the vision tower — it was making DSpark speculative decoding and decode-context-parallel run together, which the stock stack refuses. Decode kept, KV doubled, images cost about a second.
Serving GLM-5.3-Flash on 4× DGX Spark: the attention route is the whole problem
GLM-5.3-Flash looked blocked on GB10: the sparse-MLA kernel path wanted a rope dimension the model does not have, and every workaround wedged the GPU. The fix was to stop patching that path and take a different one. Decode is now flat from 800 to 93K tokens of context.
Serving DeepSeek-V4-Flash on 4x DGX Spark: the 3.76.1 recipe
An in-house vLLM image serving DeepSeek-V4-Flash at ~71 tok/s decode, 400K context, thinking enabled by default — distilled from a month of benchmarking, three confirmed bugs, and one configuration audit. Extracted verbatim from production.
Runtime guides
3GLM-5.2
Per-component mixed precision — FP8 attention, NVFP4 dense/shared experts, MXFP4 routed experts — so GLM-5.2 fits and runs efficiently across a 4-node DGX Spark GB10 cluster.
GLM-5.3-Flash
Serves GLM-5.3-Flash on GB10 by routing attention through the SM90 NoPE MLA backend extended to compute capability 12 with FA2, rather than fighting the SM120 sparse-MLA path that assumes a rope dimension this model does not have.
DiffusionGemma 26B-A4B-it
~49GB → ~26GB INT8 quantization of a 26B MoE block-diffusion model, chosen specifically to run on Ampere GPUs rather than requiring the Hopper/Ada hardware FP8 needs.