THE LAB / REPRODUCIBLE RESULTS

Measured, not assumed.

Hardware, workload and cache state belong beside every performance claim.

GPURTX 3060 · 12 GB

CPURyzen 5 5500

MEMORY32 GB DDR4

ENGINE / QUANTStrata · IQ1_M

CURRENT SOURCE BRANCHES / ALPHA

Windows 0.4.0-alpha.6 ↗Linux 0.1.0-alpha.2 ↗

Last verified 2026-10-28. Source versions are not automatically published release tags.

OCTOBER 26, 2026 / SHORT CODING REQUESTS

Cache state changes the picture.

Download raw JSON ↓

Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF · Temperature 0 · Reasoning off · Maximum output 256 tokens

Warmup, uncached

17.4 tok/s

Repeat 1 · monitor off

31.8 tok/s

Repeat 2 · monitor on

35.7 tok/s

Repeat 3 · monitor off

41.8 tok/s

Repeat 4 · monitor on

45.1 tok/s

Repeat 5 · monitor off

46 tok/s

Repeat 6 · monitor on

44.5 tok/s

The 46.0 tok/s sample ended at the output limit. These are separate short runs on a warm engine, not sustained throughput or a claim that monitoring caused a speed increase.

Warm engine; one warmup, six alternating identical short coding requests. Temperature 0, reasoning off, max 256. Light samples every five seconds; process/GPU snapshots not enabled. The uncached warmup produced 237 output tokens; completed cached samples produced 243. Hardware fit and performance vary.

What we report

We distinguish prompt processing, generation speed, time to first token, total task time and cache state. A fast repeat of a short request does not establish performance on an entire repository.

What comes next

Future comparisons should hold model, quantization, prompt, output budget, background load and runtime settings constant, then repeat each condition. Longer coding tasks need correctness checks alongside speed.