MLX vs CUDA: Qwen 3.8 Flash Next optimization duel delivers 55%+ speedups on both sides
gajesh · x · 2026-09-11
- Alibaba's Qwen team launched Qwen 3.8-Flash-Next on both the Spark and MLX communities as an optimization challenge, tracking both communities' progress on the same graph for the first time.
- The task: make Qwen 3.8 Flash Next run faster on two concrete setups — a C/CUDA engine ported from antirez's ds4 (with Unsloth's GGUF weights on a single NVIDIA DGX Spark) versus a Swift/Metal implementation.
- Both challenges have already delivered over 55% speed improvement, with CUDA currently holding a slight edge. The author frames it as friendly competition between two highly overlapping communities pushing the same model as far as possible.
Related event: Qwen 3.8 Flash Next Optimization Challenge: MLX and CUDA Both Gain Over 55%(2 posts)→
More from Infra
- EdgeRock open-sourced: 72 MiB KV cache at 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- RTX 5080 price jumps from $1658 to $2318 in ten days on Amazon — Frizzy-MacDrizzle · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Neon rebuilds its docs search entirely on its own stack: Lakebase Search, Functions, AI Gateway — matei_zaharia · 2026-10-27
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27