Qwen3.8 Max: Qwen3.8-27B Local Quantized Model: Reasoning Effort Level Test

Summary

The author tested Qwen3.8-27B on four machines: MLX 4-bit on an M5 Max, and unsloth/Qwen3.8-27B-NVFP4 running through vLLM on a DGX Spark. He observed a marked jump from thinking off to effort=low, but on the 4-bit model, xhigh can take an extreme amount of time or be truncated because of long reasoning and quantization error. Follow-up replies also provided information about a 64K budget, an 8-bit retest, and memory consumption.

Key facts

• Results were averaged across two runs, with each run lasting several hours.

• effort=low trades higher token costs for results approaching frontier-model quality.

• On the 4-bit model, xhigh may exceed 50K thinking tokens without completing; the author used geometric means to exclude extreme values.

• The author later said that a 64K budget allowed some xhigh tasks to complete; some tests took about an hour but still reached 100%.

• The M5 Max has 128GB of RAM; the 4-bit tests used about 23–24GB, while 8-bit used about 40GB; longer contexts require additional memory.

Assessment

This is not a direct evaluation of the Qwen3.8-Max cloud model, but a local quantized test of an open-weights 27B model from the same family. Its implications for prompting and runtime configuration are to establish cost and latency baselines with low/medium first, then raise effort separately for difficult tasks, with explicit token and time limits for xhigh.

Article text

Qwen3.8-27B - Interesting test results on effort I ran tests on 4 machines overnight and have been tweaking the tests al… This is a necessary excerpt; read the original source for full context.

Thread additions

I ran the model with a 64k token budget and eventually had to use that for xhigh to complete the tasks that were failing… This is a necessary excerpt; read the original source for full context.