Qwen3.8-27B Benchmarks: 672 tps on RTX 3090

Local benchmarks of Qwen3.8-27B have poured in since its release: SGLang offers Day-0 support, and with NVFP4 quantization plus DSpark speculative decoding, a single RTX 5090 decodes at 206.1 tok/s while the RTX 6000 Pro (96GB) hits 200-223 tok/s in single-stream throughput; the RTX 4090 delivers 73 tok/s for coding with a 128k context, which testers deem fully usable. A 27B-class model now runs smoothly on a single consumer GPU.

Confirmed

• Relayed by @StefanoGogioso: SGLang provides Day-0 support for Qwen3.8-27B; an RTX 5090 with NVFP4 and DSpark decodes at 206.1 tok/s, and 38.28 tok/s on DGX Spark
• Relayed by @BanghuaZ: an SGLang script on the RTX 6000 Pro (96GB) using NVFP4 weights plus DSpark speculative decoding reaches 200-223 tok/s single-stream throughput with native long-context support (post summary truncated; exact figures not fully shown)
• @iamMess pushed inference to 672 tps on an RTX 3090 through extreme optimizations: W4A16 quantization, FP8 KV Cache, and converting lmhead and embedtokens to int8
• @cocktailpeanut relayed user zast57: with Pinokio one-click deployment of Ollama + OpenWebUI, an RTX 4090 runs the Qwen 3 series 27B at nearly 100 token/s for proofreading and other text tasks
• @julianharris tested local coding on an RTX 4090 with unsloth 4-bit quantization: 73 tok/s at 128k context, fully usable; disabling thinking mode speeds things up to 78 tok/s but quality drops noticeably

Unconfirmed

• @iamMess's 672 tps does not specify whether it is decode or prefill speed; it sits far above other cards' decode rates ranging from the tens to the low 200s, and the measurement methodology awaits clarification from the author
• Model names vary across posts (e.g., Qwen3.6-28B, Qwen 3.28:27b); this roundup standardizes on Qwen3.8-27B

Why it matters

• Locally runnable at launch: SGLang Day-0 support, NVFP4, speculative decoding, unsloth quantization, and Pinokio one-click deployment landed almost simultaneously, dramatically shortening the cycle from new model release to consumer-GPU usability
• 128k long-context coding still hits 73 tok/s on a single 4090, so mid-range GPUs can already shoulder everyday coding-assistant workloads
• The thinking-mode trade-off (a small speedup when disabled but noticeably worse quality) gives local-deployment users a direct configuration reference

2026-08-16 ~ 2026-08-17 · 6 related posts

• Qwen 27B hits nearly 100 tokens/s on a local RTX 4090 via Ollama and Pinokio — cocktailpeanut · 2026-08-16
• Qwen3.8-27B hits 223 tok/s on RTX 6000 Pro with NVFP4 — BanghuaZ · 2026-08-17
• [source] Qwen 3.8 27b Local Coding Test: 73 tok/s at 128k Context, Usable Performance — julianharris · 2026-08-17
• [source] RTX 3090 runs Qwen3.8-27B at 672 tps via extreme quantization — iamMess · 2026-08-17
• Qwen3.8-27B hits 206 tok/s on single RTX 5090 via SGLang — StefanoGogioso · 2026-08-17
• [source] Qwen3.8 27B Hits 39 tok/s on RTX 4080 Super, Good for Local Coding — BopSupreme · 2026-08-17