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++
# Tuning through the App Host HTTP API
# repo: 79ead84224f6b6d982b9ca71b02f6a4174cb55e3
# The job contract is Strata Control's (strata_control/tuning.py), driven through 'strata tune'.
# The plan is computed from the SELECTED ENGINE BINARY's own --help output, so a sweep cannot
# be offered a dimension the engine would refuse.
POST /api/tune/plan {engine, backend, model, objective?, trials?, repeats?}
GET /api/tune/jobs recorded jobs, newest state
GET /api/tune/job?id=<id> one job's durable record and report
# Observed, metal backend against the real built engine:
# searches: ple_io, kv, kv_grow, spec_min_p
# excluded: 7 dimensions, each with a reason
# flags read from the binary: true
# Observed, cuda backend against the SAME (Metal) binary:
# vram_reserve -> the engine binary does not accept --vram-reserve-mib
# pcie_frac -> the engine binary does not accept --pcie-frac
# (the remaining CUDA dimensions are offered because this binary does accept their flags,
# which is the check doing its job rather than trusting the backend label)
# There is deliberately no run route: a sweep needs a measurement harness that boots the
# engine per candidate, and a fabricated measurement is worse than none.