# Disk Is the New VRAM: The Expert-Streaming Trend Colibrì Belongs To

- Published: 15 September 2026
- Updated: 15 September 2026
- Author: RenderBob team
- Category: Technical
- Canonical: https://renderbob.io/blog/disk-streaming-expert-offloading-trend

Colibrì is the visible example of a wider pattern: tools that treat fast NVMe as a planned inference memory tier, not swap of last resort.

Colibrì is the most visible example of a pattern bigger than one project: a small, growing cluster of tools treating fast NVMe storage as a deliberate tier in the inference memory hierarchy. Disk is a planned destination for data that is cheaper to store than to hold resident, not a slow fallback you fall into when RAM runs out. Community discussion around Colibrì points to at least two adjacent projects working the same territory from different angles: flash-moe, reported getting 5+ tokens per second on Apple's M3 Max through a different implementation, and hypura, working general-purpose SSD streaming rather than one model family. Add the established expert-offloading work in projects like kTransformers, and a category is visible: MoE-aware storage tiering as its own discipline.

The technical insight underneath all of them is the same one this blog's memory-hierarchy coverage explained in the VRAM, RAM and disk context: not every part of a large model needs to be resident at once, and the parts that do not can be placed rather than held, staged from a slower tier exactly when a router proves they are needed. What changed in 2026 is the confidence to lean on this deliberately for models previously assumed to require data-centre-class VRAM. Colibrì's own framing, "a JIT, but for weights," captures it: a compiler JIT does not compile the whole program; it watches what actually runs and compiles the hot paths just in time. These engines make the same bet about a model with hundreds of billions of parameters. Most of it is cold for any given token, so most of it does not need to be resident.

A motion-graphics studio whose actual workloads are diffusion video models rather than MoE language models still has two reasons to care. First, the underlying technique (measured routing heat, learned caching, prefetch that hides staging latency, striping reads across multiple drives) is architecture-general, and diffusion and MoE video models increasingly borrow ideas across categories, including the hybrid attention and DeltaNet-style mechanisms already covered on this blog. Storage-tier thinking is likely to migrate the same way speculative decoding and quantization already have. Second, this trend is concrete evidence for the "extract more from hardware you own" strategy this blog argues for in the GPU shortage. If an LLM community can run a 744B model on 25GB of RAM and no GPU by rethinking memory placement, scarce fast memory is a placement problem, not a hard ceiling, for a studio wondering whether it truly needs another $5,000 GPU.

The caveat, consistent across every project in this space: speed is the variable that suffers, and it suffers a lot, when fast memory is scarce. That is a real trade-off. For workloads where throughput matters less than access at all, it is a trade-off worth understanding rather than dismissing.