5-Hour Local Coding Agent Session on Qwen3.8-27B: 97 tok/s Streaming, 70% Context Compaction

illgettheownerforyou · reddit · 2026-08-18

The author ran a real 5.32-hour multi-slice project development session with Qwen3.8-27B loaded locally via the NInfer inference engine plus the official DeepSeek Harness (Code Mode), on an RTX PRO 6000 Blackwell (95.59 GiB VRAM), Ryzen 9 9950X3D, and WSL2. The longest single autonomous turn ran 1.06 hours and completed normally.

Key numbers:

The standout result is automatic context compaction: DSH compacted twice within the same long turn — 201,785 → 83,285 tokens (58.7% reduction) and 208,730 → 64,159 (69.3% reduction), each taking 149 seconds, after which the agent retained implementation state and kept working. The author calls it the first local setup where genuinely long agent sessions feel practical instead of collapsing under their own context.

Note: since the unsloth NVFP4 safetensors couldn't load in NInfer at setup time, measurements used NInfer's internally built Qwen3.8-27B quantized artifact.

Related event: Qwen3.8-27B local deployment tested: stable single-GPU agent runs and 128K context(3 posts)→

Original post →

More from coding & agent

coding & agent channel →