Reddit · 多媒体与视觉
Qwen 3.8 27b with DSH(DeepSeek Harness) is Amazing!! Experiences so far and perfomance.
The author reports testing Qwen 3.8 27B with DeepSeek Harness for about 10 hours and roughly 10 million input tokens. They observed automatic context compression near 90K context without losing the task, report about 37 tokens/s on an RTX 3090 as context grows, and describe using UD Q4 K XL plus vision F16, 92K context, MTP and ngram. The author identifies…
SOURCE-DECLARED INSTALL
Open the public source page ↗MEDIA REFERENCES
Captured in public view
CONTEXT
Why it is here
The author reports testing Qwen 3.8 27B with DeepSeek Harness for about 10 hours and roughly 10 million input tokens. They observed automatic context compression near 90K context without losing the task, report about 37 tokens/s on an RTX 3090 as context grows, and describe using UD Q4 K XL plus vision F16, 92K context, MTP and ngram. The author identifies…
REFERENCES
Linked evidence
Evidence updated 2026-08-16T11:57:12Z. Interaction numbers are platform-native snapshots; the evidence panel records the metric source and observation time. NULL means the public page did not expose a number at collection time.