Qwen 3.8 27B review and vLLM setup with Hermes

Qwen 3.8 27B powers an arcade coding test through Hermes Agent in this local LLM review. The speaker reports that the JavaScript project took 2 hours and 40 minutes to generate from one prompt that also requested a full review. The resulting games have sound and score tracking, but testing exposes awkward controls, a pause key that resets a game, and unreliable audio toggles.

The speaker judges the code and tool use better than Qwen 3.6 27B in earlier Hermes sessions, reporting no failed tool calls in this run and more successful use of Playwright. These are observations from one project, rather than a controlled comparison. The test uses a fresh Hermes installation in Proxmox.

The setup walkthrough covers vLLM with tensor parallelism set to four, thinking enabled, automatic model length, and a reported 262,144-token context. The speaker sets GPU memory utilization to 0.93 and recommends disabling MTP speculative decoding after finding its speed benefit unconvincing for this workload. The host contains four RTX 3090s and one RTX 4090 on a WRX80 Threadripper platform; the inference configuration selects four GPUs.

A later skill-generation task reaches about 41.2 generation tokens per second. Suggestions about fitting FP8 or NVFP4 variants on smaller GPU configurations remain tentative, and the speaker leaves their quality tradeoffs untested.