Long agent
NVFP4 · SGLang · FP8 E4M3 KV · FlashInfer · MTP off
~69.3 tok/s
steady decode median; ~60.8 at 80K+Single RTX 5090 · 128K context · one runtime at a time
Qwen3.8-27B dual-runtime agent recipe
Use SGLang NVFP4 for long-horizon work and llama.cpp Q5_K_M with MTP3 for bounded implementation. The router selects a role before loading a model.
No model weights are hosted here. Download artifacts from their original Hugging Face repositories and retain their licenses.
Bounded worker
Q5_K_M · llama.cpp · Q8 K/V · MTP3
151.72 tok/s
short decode; 98.59 at 80KRaw decode TPS is not agent throughput
In one counted autonomous web-project trajectory, the slower SGLang route completed the task in 526.388 seconds. The Q5 route had made no semantic edit after at least 541 seconds. Q5 still passed 10/10 bounded TSX fixtures and a multi-file build/test task.
| Runtime | Bounded coding | Long autonomous |
|---|---|---|
| SGLang NVFP4 | runtime/tool gates pass | completed with two acceptance caveats |
| Q5 + MTP3 | 10/10 + multi-file pass | N=1 trajectory regression |