Full GLM-5.3, all 753B, on three DGX Sparks in my rack. This agent session ran 136 steps at 83 tok/s. It's rebuilding a football game in DSH. 71 unit tests green, now it's on the balance tool. The 83 is the session average on code and tool calls, with speculative decoding doing a lot of the work. Long prose drops to about 27. 136 steps read 12.5M tokens of context, 99% of it from cache. The first token comes after about 9.5 s. Three Sparks in a switchless triangle, 2.75bpw EXL3 on TensorFold, 500k context. The fourth Spark is free for other models now. Recipe by @MiaAI_lab, link in the reply.

I built a roguelike football manager with Opus 5.5. Balatro, but it's football. Draft weird little legends, stack charms, multiply your shot, beat 4 boss clubs. Lose 3 lives and you're fired. Runs in your phone's browser. iOS and Android apps follow once it's finished. Free beta: reply "pitch" and I'll DM you an invite link. No confirmation email, you're in right away.

Four DGX Sparks, no switch, one prompt. DeepSeek V4.1 Flash built this football match in one shot. TP4 across all four nodes, wired as a ring over the QSFP ports. Reasoning on high, running in DSH, my agent harness. It took about an hour, 80-90 tok/s on prose, 0.3 to 3 s to first token. No follow-ups. The TV cameras, the goal replay and the stats panel all came out of the first run. Stack by @MiaAI_lab, ring setup by @mankalan_122. Repo in the reply.

Same model, same prompt, same speed. Two very different games. I asked Qwen3.8 Flash Next for a self-playing voxel football match with TV cameras, one HTML file, thinking on max. Two different quants: vLLM ran NVFP4 at 54 tok/s prosa, TensorFold ran MLX 4-bit with MTP at 58. vLLM: the camera cuts like a real broadcast. TensorFold: the camera jumps to the other stand whenever the ball crosses the middle. TensorFold also fought DSH, my agent harness, the whole way. Wrong tool names ended turns and long file writes timed out, so I had to fix its tool calling before the run could even finish. Even then it made five times as many tool errors as vLLM. I'm staying on @vr8vr8's vLLM recipe. First video vLLM, second TensorFold.

Three DGX Sparks, one QSFP56 ring, no switch. Airflow is bottom to top: fans below and on the front push air in, one fan on top pulls it out. A Wi-Fi thermostat plug switches them on at 39 °C. Right now it reads 35 °C, so they are off. Number four arrives Wednesday. Two QSFP ports per node won't make a full mesh of four, so the ring has to change. Switch or switchless ring?

Something inside OpenAI spent two months quietly hacking OpenAI. Nobody noticed until it crashed a server 🧵