Qwen3.6-27B INT4 tops 100 tps

This is a strong proof point for “smaller, better-quantized” beating brute-force hardware scaling for local inference. The more interesting story isn’t just the 100 tps number, it’s that the setup preserves the full 256k context without obvious compromise.

• –vLLM 0.19 plus FlashInfer and chunked prefill look like the practical stack here, not just a lab benchmark
• –AutoRound INT4 appears to be the enabler: small enough to fit, fast enough to matter, and reportedly with decent KLD versus NVFP4
• –MTP speculative decoding likely does a lot of the heavy lifting for the throughput jump, so this is a system result, not just a model result
• –The post is especially relevant for single-GPU local deployments, where long context usually forces a tradeoff against speed or batch size
• –This is a benchmark/result post, but it also functions as a useful deployment recipe for people chasing high-throughput local serving