Introducing Full GLM-5.3 on 4x RTX PRO 6000: recipe 1.0.0 with TensorFold v0.6.5 🔥 4x RTX PRO 6000 (250 W per GPU): - 92 tok/s on prose, single stream - 140 tok/s on prose, 5 concurrent streams - 195 tok/s on structured output, 5 concurrent streams - Prefill up to 2,081 tok/s - Next turn in a 130K-token conversation: 4.14 s to first token - Up to 5 concurrent requests, 1M context What's new: - TensorFold v0.6.5 with 151 engine patches: 106 ported from the original full GLM-5.3 recipe, 45 of ours - Five concurrent sessions with context parallelism over one 1,245,184-token FP8 KV pool - RAM and NVMe prompt tiers: a 747,292-token prompt resumes in 4.07 s from RAM - Teacher KL 0.0468 with FP8 dense and FP8 KV, lower than the 0.0568 of our earlier build - One command: ./start.sh builds, downloads and serves Special thanks to @MiaAI_lab for the EXL3 checkpoint and base patches, and to @ashxhart for TensorFold!

GLM 5.3 Flash on 2x RTX PRO 6000 with TensorFold is out! 🚀🚀 Our one-command recipe for two RTX PRO 6000 GPUs. 2x RTX PRO 6000 (300 W per GPU): - 238 tok/s on prose, single stream - 449 tok/s on prose, 4 concurrent requests - 651 tok/s on structured output, 4 concurrent requests - Prefill up to 6,908 tok/s - Next turn in a 130K-token conversation: 0.51 s to first token - Up to 4 concurrent requests, 512K context What's new: - A recipe for 2x RTX PRO 6000: one host, half the GPUs of our 4x recipe - 512K context per request, two full 512K sessions at once (1.06M-token shared pool) - Vision on by default - One-time TP2 launch-table autotune - TensorFold v0.6.5 with the patch series from Mia's AI Lab and Aevonix Research, and the same exactness gates - One command: ./start.sh builds, downloads, compiles and serves Special thanks to @MiaAI_lab for the EXL3 quant, and to @ashxhart for TensorFold!

To EVERYONE who says local is useless. This is the reason you are wrong and will become more wrong as time goes by. Local is Sovereignty, Local is Freedom It is the ONLY way to guarantee you have equal access to intelligence

I’m building “The All Spark” a 36x DGX Spark cluster to run my own local agents and support the community with compute. 24 Sparks running today on a single cluster with the rest coming online after a quick power upgrade to the house 😅

MiMo V2.6 Pro RL on 8 DGX Sparks: 17.8 → 68.3 tokens/s with DFlash. 3.8× end-to-end speedup across four long structured-output tests, versus the same runtime without speculation. We’ve released the patches, setup and raw results: