Hibana: Fast Agentic Inference on Unified-Memory Devices A Cornell University-led research team published Hibana on November 4, 2026, a co-designed model, memory and decoding system that runs many local agents on a single DGX Spark unified-memory device. Hibana uses 3-bit weights with an 8-entry FP8 codebook (3.25 bits per weight including an 8-bit scale per 32 weights), INT8 recurrent state, a 2-bit draft head that shortlists 16 of over 100,000 tokens and agrees with the FP8 head on 99.76% of tokens, and SSD spilling of agent state; on Qwen3.8-Flash-Next with 192 conversations the SSD tier lifts cache hits to 80–98% and cuts time to first token by up to 42× against vLLM, llama.cpp and TensorFold baselines. The work targets the 128 GB shared memory pool and limited bandwidth of unified-memory devices, where every generated token streams the weights. Paper · November 4, 2026 On a unified-memory device, the model and every agent's history share one memory. Hibana co-designs the model, its memory, and its decoding so many agents run fast on a single DGX Spark.