WHAT NEW IN X10.30 FRI

AI collections from X · 22 collections · 60,637 posts

Back to No.020 DeepSeek V4.1 Flash hands-on tests and comparisonsUpdated 12 hours ago
deepseek-v4.1IMAGE

DeepSeek V4.1 Flash on 4x DGX Spark: faster overall, 16-user throughput down ~15%

Running V4.1 Flash on four DGX Sparks with TensorFold v0.3.5 and 4.0bpw weights plus two patches improved the fleet, at the cost of about 15% lower 16-user throughput.

Tech2Wild
@Tech2WildArchived 2 days agoOriginally posted 2026.10.27
DeepSeek V4.1 Flash on 4x DGX Spark: faster overall, 16-user throughput down ~15%
Tech2Wild @Tech2WildOriginally posted 2026.10.27
Open original on X

Before vs now on our 4x DGX Spark DeepSeek V4.1 Flash fleet. Full credit to @bertholomusai for the TensorFold v0.3.5 recipe, engine and 4.0bpw weights. We added 2 small patches, both RoCE lanes and a 29.4M KV pool. One tradeoff: 16-user throughput is down ~15%.

x.com/tech2wild/status/2114866354612228293Archived 2026.10.27 12:08
Open original on X