qwen38

Episodes (3)

Trajectory notes

• 2026-09-06T21:07:17Z: qwen38-27b-16gb-quant-benchmark closed (faded) — The radar already tracks this model’s local-agent performance under “qwen38-27b-local-agent-capability,” while Scott’s “gamepc — self-hosted GPU model zoo” and “Hardware-aware local inference” make quantization-versus-VRAM eviden
• 2026-09-03T06:26:57Z: qwen38-max-open-weight-coding closed (faded) — Scott already holds the decisive position in “Usable Mass Over Unusable Power”: nominally powerful open weights matter only when serving cost, latency, quantization fidelity, and harness performance make them deployable. This new r
• 2026-08-19T00:26:32Z: qwen38-27b-24gb-long-context-throughput closed (faded) — If independently reproduced, the result would extend Scott’s hardware-aware local-inference work by showing that a quantized 27B model can combine a 256K resident context with interactive MTP throughput on a constrained 2