I trained an adapter on Qwen3.8-27B that speed up 1.71x prefill (ref kis LLKVApprox took the learning from DeepSeek-V4.1-Flash CED to qwen series)

3,396 tok/s prefill (1.71x @ 6,603 tok, 69.8% greedy match vs baseline untrained projector) - 147 tok/s decode single stream

Sean PhanSep 133

Devlog #5 Nanoharness First larger one shot ~10 minutes.

I might be biased, but this is one of the better results. Model: Deepseek V4.1 Flash Terrible token usage so far, room for improvement: in: 345.957 out: 255.642 cached: 22.521.728 cache hit: 98% reasoning: 103.304

balega_devSep 13

On Deepseek Harness, quite usable, enjoyed..

Erman ErogluSep 13

Shipped a full drone-combat game in ONE html file procedural terrain, forests, wildlife, a 16-plane swarm, guided missiles.

Chase + FPV cams, HUD, explosions. Built end-to-end by the DSH and huihui_qwen38_27b. 🚁💥 100 t/s average decode speed, 26 M tokens in total, all for

Erman ErogluSep 131