Qwen3.6 128K at 30 t/s on RTX 5080

On an RTX 5080 16GB, Ryzen 9700X, 96GB DDR5, Windows 11 setup, Qwen3.6-35B-A3B was served locally via llama.cpp to build a long-context coding workflow for Claude Code.

• Claude Code was connected via the Anthropic-compatible /v1/messages endpoint, with the goal of stably maintaining 65K–128K context.
• The fork used was craftogrammer/llama.cpp-adaptive-turboquant from the Madreag/turbo3-cuda lineage, applying TurboQuant/TCQ to the KV cache.
• CUDA 12.9.1 was required; CUDA 13.x produced broken output, and 13.1 caused segfaults in the MMQ kernel.

The result was 30 t/s at 128K context and 89 t/s for new tokens. The author reported no perceptible drop in coding quality even at long context.