Matching stories Newest first
Community user benchmarks Qwen3.8 Flash-Next and 27B long-context throughput on four Intel B65 GPUs
Throughput sweeps from 1 to 16 concurrent requests across context lengths up to 200K show Flash-Next W4A16 reaching 84.5 tokens/sec at one request with MTP-3 enabled. Numbers are overall output throughput, not prompt-processing speed, and starred results include queueing.