Back
RCreddit.com
12
·2 days ago·RSS
Not on the current live radar

Qwen3.8-Flash-Next optimised for Macs

View original
Plans & limits

Heat trend

Collecting trend data

The percentage is based on available heat signal, not comment count or independent people.

AI summary

Qwen3.8-Flash-Next has been optimized for Macs, with further optimizations for devices with limited RAM and cache. While disabling MTP (Multi-Tensor Processing) can achieve higher prefill speeds (180-190 tps at 4K context), enabling MTP significantly boosts decode performance by 70%, reaching 22 btps, despite slightly reducing prefill to 170 tps (at 4K) or 150 tps (at 256K context) due to increased RAM usage for KV cache. MTP is generally recommended for its overall performance benefits.