Qwen 3.6 context window is 32k, OpenRouter reports 262k for the same model

[Evrard-Nil]

Source: @PierreLeGuen in #private-inference, 2026-05-20 18:09 CEST.

Problem

Our deployed Qwen3.6-35B-A3B-FP8 advertises a 32k context window. OpenRouter lists the same upstream model at 262k: https://openrouter.ai/qwen/qwen3.6-35b-a3b

Users on our endpoint get a much smaller usable context than competing providers for the same model.

Investigation

The current Qwen 3.6 config (cvm-compose-files, deployed v0.0.163 with sglang) is set to 32k somewhere — likely a sglang/vllm --max-model-len arg in the compose file, or a kv-cache/VRAM constraint forcing it down.

Tasks

• Confirm where the 32k limit is set (compose file, sglang flag, KV-cache budget)
• Decide target context window — 262k would need significantly more KV-cache GPU memory; intermediate values (128k, 64k) may be a reasonable trade-off
• Bump and redeploy
• Update /v1/models advertised context_length

cc @PierreLeGuen — flagged this in #private-inference; want to confirm whether it's already on your plate or if this issue is the right home for it.