GPU kernel bench leaks: DeepSeek v4.1 flash beats GLM-5.3 and Haiku 5.5
teortaxesTex · x · 2026-10-10
A user shared GPU kernel benchmark results claiming DeepSeek v4.1 flash outperforms GLM-5.3 flash and Haiku 5.5, hy4 preview matches Gemini 3.8 flash, and Qwen 27B FP8 beats a Q4 Strata Next version. Reposter teortaxesTex notes this is unsurprising given DeepSeek's focus, possibly tied to its migration toward Ascend chips where GPU kernel optimization is a headache. These version numbers are unverified.
More from Infra
- Microsoft's open-source BitNet runs a 100B LLM on a single CPU with 1.58-bit weights — JafarNajafov · 2026-10-10
- Open-source typesafe-computer-use cuts agent cost 155x by skipping screenshots and frontier models — thisguyknowsai · 2026-10-10
- Why RAM prices went up 500%: a look at the memory price surge — airesearch12 · 2026-10-10
- Colocating Compute With Stranded Renewables Could Unlock Curtailed Power — AccBalanced · 2026-10-10
- Micron didn't end up with zero Rubin HBM, as analyst commentary backpedals — suchenzang · 2026-10-10
- Memory now eats 40% of PC costs as AI demand drives shipments down 20% — APPSO · 2026-10-10