0.3

New model families (BitNet 1.58-bit, PLaMo 2, Apertus, Seed-OSS, dots.llm1, Falcon-H1, LFM2/LFM2-MoE, IBM Granite) and audio I/O (TTS speech + Whisper ASR endpoints). Fused decode-MoE kernel wired across MoE families, Turbo KV cache modes with a per-family advisor, CUDA (GB10) warmup/JIT fixes, bf16 and mixed-bit quant correctness, and decode-performance work for M1 through M4.