Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
// include/strata/core/draft_cap.hpp - the widest verify window --draft dflash2 uses on each card class (measured A/Bs, notes-dflash2-p1.md; recommend, never force).
#pragma once
#include <cstddef>
namespace strata::core {
/// 0 = no cap (the --spec given stays). `uma`: Strix Halo / iGPU (gfx1151); `cc_major`: the CUDA compute capability's major (0 on HIP).
/// - Strix Halo / UMA: 5 (agent-safe; 6 ties on agent prompts)
/// - Blackwell (sm_120, cc 12): 6 (+1.9% at 6 on an RTX PRO 5000, 8 was worse)
/// - everything else: unmeasured, --spec as given
inline int dflash2_auto_cap(bool uma, int cc_major) {
if (uma) return 5;
if (cc_major == 12) return 6;
return 0;
}
/// A 12 GB NVIDIA card (< 13 GiB, CUDA): the MTP drafter was faster in most measurements, which the engine says once.
inline bool dflash2_small_nvidia(bool cuda, size_t total_bytes) { return cuda && total_bytes < ((size_t) 13 << 30); }
} // namespace strata::core