Qwen3.8-27B local benchmarks: from RTX 4080 Super to RTX 6000 Pro
Following the release of Qwen3.8-27B, the community ran extensive local benchmarks across GPUs from the RTX 3090 to the RTX 6000 Pro. High-end cards paired with SGLang and quantization can exceed 200 tok/s, while mainstream consumer GPUs (4090, 4080 Super) remain viable for long-context coding and everyday generation, indicating broad feasibility for local deployment.
Confirmed
- @iamMess reached 672 tps on an RTX 3090 via W4A16 quantization, FP8 KV cache, and int8 lmhead/embedtokens
- @StefanoGogioso relayed SGLang's Day-0 support: 206.1 tok/s decode on a single RTX 5090 with NVFP4 and DSpark speculative decoding; 38.28 tok/s on DGX Spark
- @BanghuaZ relayed: RTX 6000 Pro (96GB) achieves 200-223 tok/s single-stream with SGLang, NVFP4 weights and DSpark, supporting native 256k context
- @rohanpaulai shared a demo at 115 tok/s on RTX 5090 (32GB VRAM); official BF16 weights are 55.6 GB
- @julianharris measured 73 tok/s on RTX 4090 with unsloth 4bit at 128k context for coding; disabling thinking mode raises it to 78 tok/s with notably lower quality
- @BopSupreme reported 39 tok/s on RTX 4080 Super (16GB), TTFT 200-230ms, 34 tok/s under a 26K-token real workload
- @cocktailpeanut relayed user zast57: RTX 4090 via Pinokio one-click Ollama + OpenWebUI runs the 27B model near 100 tok/s on proofreading tasks
Unconfirmed
- @iamMess's 672 tps figure does not clarify prefill vs. decode, a large gap versus other measurements
Why it matters
- Extreme optimization cases and high-end results show the hardware ceiling, while smooth mid-range performance proves broad local deployability
- The thinking-mode speed/quality trade-off and the 115 vs. 206 tok/s comparison on 5090 offer concrete guidance for deployment choices
2026-08-16 ~ 2026-08-17 · 7 related posts
- Episode 1: Developers Test Run 2.4T Parameter Qwen Model on Consumer Hardware(2026-08-14, 3 posts)
- Episode 2: Qwen3.8-27B Benchmarking on RTX 6000(2026-08-15, 2 posts)
- Episode 3: Community Shares Qwen3.8-27B Deployment on 32GB VRAM(2026-08-15, 4 posts)
- Episode 4: RTX 3090 Optimized for Efficient Qwen3.8-27B Inference(2026-08-15, 3 posts)
- Episode 5: Qwen3.8-27B Hits 50 tok/s on Dual RTX 3060s(2026-08-15, 2 posts)
- Episode 6: Qwen2.5 Local Performance Review(2026-08-15, 2 posts)
- Episode 7: Qwen3.8-27B local benchmarks: from RTX 4080 Super to RTX 6000 Pro(2026-08-16, 7 posts)
- Episode 8: Running Local Agent Models on Mac: Memory Is the Deciding Factor(2026-08-16, 2 posts)
- Episode 9: RTX 3090 users seek best local setup for running Qwen3 models(2026-08-16, 2 posts)
- Episode 10: Qwen3.8-27B local deployment tested: stable single-GPU agent runs and 128K context(2026-08-16, 3 posts)
- Episode 11: Qwen3.8-27B Inference Optimized for RTX 3090(2026-08-18, 2 posts)
- Episode 12: Inco AI Launches DFlash 2, Speeding Up LLM Inference Up to 4.6x(2026-08-19, 4 posts)
- Episode 13: llama.cpp's New DFlash2 Boosts Qwen Inference Speed Up to 3x(2026-08-20, 2 posts)
- Episode 14: Qwen3.8-27B Inference Hits 381 TPS on RTX 3090(2026-08-20, 2 posts)
Primary sources
- 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
- Qwen3.8 27B Hits 39 tok/s on RTX 4080 Super, Good for Local Coding — BopSupreme · 2026-08-17
- [source] RTX 5090 Demo: Running Qwen 3.8 27B Locally at 115 tok/s — rohanpaul_ai · 2026-08-17