Hands-on: V4.1 Flash EXL3 2.9bpw on dual DGX Spark as an implementation agent

The author ran a quantized V4.1 Flash on two DGX Sparks as an implementation agent with a separate reviewer, finding it useful but prone to overstating evidence.

Tom Turney

@no_stp_on_snekArchived 10 days agoOriginally posted 2026.10.22

Hands-on: V4.1 Flash EXL3 2.9bpw on dual DGX Spark as an implementation agent

Tom Turney @no_stp_on_snekOriginally posted 2026.10.22

Open original on X

1/ I've been running DeepSeek V4.1 Flash (EXL3 2.9 bpw, 2x DGX Spark, TP2) as the implementation agent on a real feature, with a separate reviewer checking its work (astra medium). It's genuinely useful, but its reports can outrun its evidence. Here's what made it work 🧵

x.com/no_stp_on_snek/status/2112997600680919155Archived 2026.10.22 12:00

Open original on X