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++
# Engine tuning: the plan for this machine's real Metal engine
# repo: 3044521c760a860df74f36304bb936a80a041fbd
# command: strata tune plan --engine build-metal/strata --backend metal --model qwen3.8-flash-next
engine /Users/nikoveit/Documents/deepseek-harness/strata-work/strata-rebuild/build-metal/strata
flags read from the binary
backend metal objective generation repeats 3
bounds at most 16 trials, 1800s total
searches:
ple_io --ple-io values ['direct', 'mmap', 'ram']
kv --kv values ['fp16', 'int8', 'q4_0', 'k8v4']
kv_grow --kv-grow values [False, True]
spec_min_p --spec-min-p values [0.001, 0.1, 0.5, 0.9]
excluded:
expert_cache the Apple Silicon engine does not parse this flag; it is an NVIDIA/AMD device concept, so a sweep could only record a boot failure
host_core the Metal build's CPU-side pool is not exposed; the flag parses but its entry point is one of the unported kernels (metal/src/engine/stubs.cpp)
pool_affinity the Metal build's CPU-side pool is not exposed; the flag parses but its entry point is one of the unported kernels (metal/src/engine/stubs.cpp)
pool_workers the Metal build's CPU-side pool is not exposed; the flag parses but its entry point is one of the unported kernels (metal/src/engine/stubs.cpp)
vram_reserve the Apple Silicon engine does not parse this flag; it is an NVIDIA/AMD device concept, so a sweep could only record a boot failure
pcie_frac the Apple Silicon engine does not parse this flag; it is an NVIDIA/AMD device concept, so a sweep could only record a boot failure
prompt_cache_tail not a metal setting
Applying a result is a separate step; a sweep never writes the active profile.
# The same command for a CUDA backend, showing the NVIDIA-only dimensions it may search:
engine /Users/nikoveit/Documents/deepseek-harness/strata-work/strata-rebuild/build-metal/strata
flags read from the binary
backend cuda objective generation repeats 3
bounds at most 16 trials, 1800s total
searches:
ple_io --ple-io values ['direct', 'mmap', 'ram']
kv --kv values ['fp16', 'int8', 'q4_0', 'k8v4']
kv_grow --kv-grow values [False, True]
expert_cache --expert-cache values [0, 1024, 2048, 4096, 8192, 12288, 16384]
host_core --host-core values ['first', 'last']
pool_affinity --pool-affinity values ['all', 'p-cores', 'auto']
pool_workers --pool-workers values [0, 2, 4, 8, 12, 16]
spec_min_p --spec-min-p values [0.001, 0.1, 0.5, 0.9]
prompt_cache_tail --prompt-cache-tail values [False, True] (benchmark only: never exported)
excluded:
vram_reserve the engine binary does not accept --vram-reserve-mib, so a candidate using this dimension could not start
pcie_frac the engine binary does not accept --pcie-frac, so a candidate using this dimension could not start
Applying a result is a separate step; a sweep never writes the active profile.