Commit History
23 days ago
-
Lower Mac streaming batch size after separate paired qualification
Verified signatureBatch2 selected on exposed data, then evaluated on 11 newly frozen recordings against original CUDA BF16. Full references, raw events and errors preserved. Timed matched captions p95 0.724s, max0.818s; no universal word-latency claim. Original model weights and CPU/GPU runtimes unchanged.52ecfff -
Release complete mixed Q3/Q4 Audio8 with portable CPU, MLX and exact GGUF GPU consumers
Verified signature1.882 GB shared CPU/MLX model; 304 MB smaller payload. Complete prospectively frozen 11-case original/CPU/MLX evaluation, portable Linux prefix parity, and actual CUDA/NVIDIA Vulkan/AMD Vulkan streamed smoke. Exact GGUF weights remain Q4_0 storage. Accuracy, memory, hard latency and AMD speed limitations remain explicit; experimental, not production ready.c49eca9 -
Publish complete experimental Audio8 Q4 model and runtimes
Verified signatureExperimental release; accuracy, memory and latency limits are documented. Release manifest SHA256: 94e528576fcb5bcabbb50295a9422f59b15515930c0298f0d22d9614c20c3a3921fd722 -
initial commit
Verified signaturef8e50fd