Got GLM-5.2 (GlmMoeDsa / DeepSeek-Sparse-Attention arch) serving on 4× GB10 over a 100G MikroTik CRS504 switch. It was quite the hassle to get DSA + MTP running, but then again Claude did 90% of the work : ) Model cyankiwi/GLM-5.2-AWQ-INT4, with a data-free 15% routed-expert prune (256→218 experts/…

also just posted more details on the build you prob could snag for speeding up your own setup~

I’m currently sitting at ~700 prefil and ~23 decode with the latest updates. If I (or Fable) can’t get DCP to work well I’d much prefer keeping the throughput.

I got the full model with DCP2 and 328k ctx working. Going to test it for a bit! The pruned model performs pertty well but seems to have some issues with instruction following on the granular level. In my claude.md I have a requirement for the model to end every generation with a phrase, and it devi…

Try madeby561/GLM-5.2-MXFP8-NVFP4-NF3-Hybrid · Hugging Face

That model is only 358B parameters and not the full 789B used in the first post.

Ah cool, if not that had been quite the REAP. ;)

[image] CosmicRaisins: I got the full model with DCP2 and 328k ctx working. Going to test it for a bit! The pruned model performs pertty well but seems to have some issues with instruction following on the granular level. In my claude.md I have a requirement for the model to end every generation…

30 tok/s? I’ve never gotten that reliably. I get ~22 in llama-bench, ~25 in agentic harnesses. 30 only with MTP bursts.

GLM-5.2 on a 4× GB10 cluster: ~22 tok/s decode, 256K ctx, Recipe

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10

tonyd615 July 7, 2026, 12:13pm 118

Just ran it and we are seeing 870k KV POOL. Currently got it set to 800k GLM-5.2 (unpruned) at 800K max context on 4× DGX Spark 20-33 tok/s — NF3-hybrid checkpoint

Topic		Replies	Views
GLM-5.2-Int4-Int8 on 8× GB10: ~1,200 t/s prefill, 33–54 t/s avg decode (generic - coding/structured) DGX Spark / GB10 gaming , llama	29	2591	August 29, 2026
GLM-5.2 on a 3× GB10 cluster: ~16,13 tok/s decode, 215K ctx FP8 TP=3 + VISION! DGX Spark / GB10 Projects	7	1521	July 26, 2026
GLM 5.2 Hybrid-FP8+NVFP4+MXFP4 + Optimal runtime recipe DGX Spark / GB10	42	4163	August 14, 2026
GLM-5.2 (unpruned) @ 200K context on 4× DGX Spark — 27 tok/s single / 52.5 tok/s @c4 DGX Spark / GB10 Projects deepseek	3	865	August 4, 2026
Followup: Mystery Solved: 4x Spark, GLM-5.2-nfp4, 24tp/s, 128k ctx, no REAP DGX Spark / GB10	34	2800	July 8, 2026
Fitting a high-quality REAP-less GLM-5.2 onto 4x DGX Spark DGX Spark / GB10 deepseek	0	2223	June 29, 2026
nvidia/GLM-5.3-Flash-FP4-Local with 2M tokens of FP8 KV Cache DGX Spark / GB10 spark , dgx	16	1268	October 20, 2026
GLM-5.3 Flash on 2× DGX Spark: 4.7M-token KV cache with fast EXL3 serving DGX Spark / GB10	6	3212	October 3, 2026
Let's optimize nvidia/GLM-5.3-Flash-NVFP4 for 2× DGX Spark DGX Spark / GB10	185	11619	October 16, 2026
GLM-5.3-Flash: 320B total parameters / 18B active DGX Spark / GB10	878	39069	October 21, 2026

GLM-5.2 on a 4× GB10 cluster: ~22 tok/s decode, 256K ctx, Recipe

Related topics