Qwen 3.8 27b with DeepSeek Harness: 10h stable, 37-60 tok/s on RTX 3090
cviperr33 · reddit · 2026-08-16
A Reddit user shares experience with Qwen 3.8 27b paired with DeepSeek Harness (DSH). On an RTX 3090, the model ran for 10 hours without errors, auto-compressed context, stayed on task, and solved problems in one shot. Speeds: 37 tok/s with large context, 50 tok/s fresh, 56-60 tok/s in Unsloth Studio. Used UD q4 k xl quant + vision F16, 92k context, MTP + ngram, 66 GPU layers. User notes default xHigh thinking mode, sometimes thinking 20 minutes. Awaits 35b MoE model as dense model needs 16GB+ VRAM.
More from Models
- 1-bit Quantized Qwen 27B Runs on 12GB VRAM at 92 tokens/s — zyxciss · 2026-08-16
- BlueBench Cyber Benchmark: GPT-5.6 Sol Leads, Kimi K3 Tops Open-Weights — vijaybolina · 2026-08-16
- GLM 5.3 Update Pace Sparks Compute Comparison with DeepSeek — teortaxesTex · 2026-08-16
- Qwen 3.8 Ablit model available on Hugging Face — Frizzy-MacDrizzle · 2026-08-16
- Qwen3.8-27B-int4-AutoRound (18GB) with working MTP spec decode — BusinessMud9586 · 2026-08-16
- MiniMax H3 Training Progress: Turbo Time LoRA Fixes Motion Lag — ostrisai · 2026-08-16