# Running Qwen3.6-35B-A3B on 2x 4090 with sglang for GitLab code review — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/11cec0703d7a54ba

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-21 14:19 ingested 2026-10-22 02:43

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xbogcj/running_qwen3635ba3b_on_2x_4090_with_sglang_for/)

## retained text



```
I’m sharing a GitLab review agent I’ve been working on. The idea was to keep code reviews on hardware I run. It’s built on LangGraph.js and uses Qwen3.6-35B-A3B (QuantTrio AWQ, 4-bit) on sglang.



The box is two 4090s over PCIe, no NVLink, TP=2. Decode is about 180 tok/s single stream at \~50k context and about 160 at \~110k. Prefill is 9-10k tok/s. At 8 concurrent requests I get 985 tok/s total.



A review run averages around 26 LLM calls and 2.3M tokens. As of now it has done 1,300+ reviews and about 70% of its comments get resolved by the author.



It sometimes catches mistakes in code written with Claude Sonnet/Opus, in the harness’s own repo and in internal projects. I believe it’s good to have different model families in the workflow, you get different perspectives from each.



Repetition is the common problem I face with local models (maybe it's the quants). Guardrails catch repeated tool calls and trailing questions, but the model looping on itself still ends about 7% of runs (56 of 847 on my dashboard).



Launch flags and the benchmark: [https://vrajpal-jhala.github.io/langgraph-harness/sglang-deployment](https://vrajpal-jhala.github.io/langgraph-harness/sglang-deployment)

The agent is MIT and also works with Ollama, OpenRouter, Gemini and Groq: [https://github.com/vrajpal-jhala/langgraph-harness](https://github.com/vrajpal-jhala/langgraph-harness)
```



document fetched 2026-10-22 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xbogcj/running_qwen3635ba3b_on_2x_4090_with_sglang_for
