DeepSeek V4.1 Flash 1017 Now Runs on Two Sparks

Two DGX Sparks are now enough for DeepSeek’s new Flash, with a 512K context.

@MiaAI_lab published a recipe for running DeepSeek V4.1 Flash 1017 on 2x DGX Sparks. It uses an EXL3-EAQv1 quant at 2.9 bpw, with a 1M total KV pool and 512K context. Prefill runs at 2000-2175 tok/s.

The quant in this recipe got about the same calibration as her GLM 5.3 quant, she says, and any other EXL3 quant can be swapped in “without any issues.” She also posted a video of four concurrent streams on the two Sparks.

@u1tra_instinct abliterated the 2x Spark build within hours: same repo and quant, with a simple ABLIT=0/1 flag.

On quality, she is still weighing it against GLM 5.3 Flash.

@MiaAI_lab says a new chart suggests her 100-HTML-file experiment on the model “was pretty accurate.” For web dev, she rates Qwen3.8-Max “much better,” more than the chart suggests, and notes she can run it at home.

She also pitted it against GLM 5.3 Flash with a prompt to build “Gummy Bears,” a single HTML file of translucent bears you can “grab, stretch, squash, throw and pour.” Both ran locally on the same harness at Max effort: GLM on her own EXL3 quant, DeepSeek on the official quant.

In replies she says physics is better on DeepSeek, and that she is “still conflicted, glm or ds.”