# Qwen3.8 27b loop testing int4 awq vs int8 — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/4d010ea2bc62f2f6

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-31 03:19 ingested 2026-11-01 02:40

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xjzm9q/qwen38_27b_loop_testing_int4_awq_vs_int8/)

## retained text



```
I know this some old news, but i got some numbers here from my heavy agent load.
In mid-October I switch from INT8 (4×3090) to AWQ INT4 (2×3090). 
Same sampling, harness(Hermes), and loop detector, so I compared about two weeks of logs from each setup.

vLLM, FP8 KV, with DFlash2.
temp 1.0 / top_p 0.95 / top_k 20, presence_penalty 0, thinking medium.
Two models: int4-awq and int8.

**Results (INT8 → INT4 AWQ)**

* Model calls: 17.7k → 11.4k
* Loops per 1k calls: **1.9 → 4.7**
* Conversations with ≥1 loop: **6% → 12%**
* Per-task success rate: **92% → 78%**
(success rate is after loop got cut, ask it continue can finish the task)

Both setups had similar reasoning lengths, so "more quant thinks longer" doesn't explain the gap.

EVERY loop was in the reasoning block, so it cooked when it cooking too hard is the problem as you guys know.

Retrying rarely fixes it. After a thinking loop cut, next call ask it continue got looped again 14–22% of the time, compared with under 1% normally. 
The context is what making loop happens, not about luck in sampling.

The INT4 gap grows with context. 
INT4 looped 1.4x as often as INT8 under ~40k tokens and 3.3x as often at 80–120k content.

With speculative decoding, loops show up as a smooth speed spike, loops ran at ~200 tok/s versus ~107 tok/s normally on my setup, because the draft model predicts repeated text very well. That makes a cheap watchdog signal.

Mostly the model saying it will act and never doing it ("Let me write it. Wait let me... How about..." 10+ line as one loop), or restating the same case analysis dozens of times.


**TL;DR:** INT8 loops too, and people in the HF discussion report the same 20k+ token thinking spirals on BF16, so this isn't purely a quant problem. INT4 or Q4 just makes it a lot worse, especially at long context above 100k~.
btw loops not that bad, just don't push model too hard, I feels on daily assistant loops happens much less.

I've saved thoes contexts that looped. When Qwen4 27b drops I can just replay them and post the new results.
```



document fetched 2026-11-01 00:38 · canonical https://reddit.com/r/LocalLLaMA/comments/1xjzm9q/qwen38_27b_loop_testing_int4_awq_vs_int8
