Deltafin fork streams full 2.8T-parameter Kimi K3 to a MacBook Pro from four SSDs

A fork of the open-source Deltafin project has demonstrated the complete Kimi K3 model — a 2.8-trillion-parameter mixture-of-experts system — running on a single MacBook Pro, with weights streamed from SSDs while the model generates. According to the benchmark published alongside the fork, the setup sustains 0.2901 tokens per second on an M1 Max laptop, or roughly one token every 3.4 seconds.

The fork, published on GitHub by Argonaut Labs, builds on the MIT-licensed Deltafin by gavamedia and adds what the authors call ARGODRIVE storage work, together with a benchmark package whose manifests and results sit in a k3-public-bench directory; the measurement tooling itself is published separately as ARGODRIVE. The Hacker News post that carried the project to the front page describes the hardware as an Apple Silicon machine streaming from four SSDs.

The weights live on disk, not in memory

Kimi K3 was built for datacenter-scale infrastructure. The project's README puts the reference deployment at 16 nodes with roughly 4.8 TB of aggregate VRAM, and the full checkpoint at about 1.7 TB on disk. No laptop can hold that in unified memory, so Deltafin treats storage as the model's home and pages in routed experts on demand. A space-saving streaming mode starts from a 215 GB footprint and caches only the experts a machine actually uses, building up a partial local copy over time.

Because K3 is a mixture-of-experts model, only a slice of the parameters participates in any given token, which is what makes disk-backed execution feasible at all. The storage layer has to be fast enough to feed whichever experts the router selects, and that is where the fork's placement and prefetching work is concentrated.

Small models guess, K3 verifies

Throughput gains come from speculative decoding rather than from trimming the model. The default setup includes Inferact's Kimi-K3-DSpark as a draft model — about 6.6 GiB on disk — which guesses ahead while the full K3 verifies every token; per the README, nothing reaches the user without the target model's approval. An optional Qwen add-on of roughly 4.3 GiB accelerates raw text completion only, and the project measured one 17-token completion running 2.7 times faster with identical output token IDs.

The README's benchmark history shows the trajectory: 0.0141 tokens per second on July 27, 2026, then 0.1311 a day later, 0.2660 on July 30, 0.2847 on August 2, and 0.2901 in the latest run — a more-than-twentyfold gain since the first recorded figure, although recent updates have shrunk to low single-digit percentage improvements.

Fidelity is the stated point

The project draws a hard line on quality. It criticizes other efforts that re-encode K3's expert bank down to around 3 bits, arguing that the resulting weights are no longer what Moonshot shipped and that nobody has measured what the compression costs. Deltafin keeps the complete 16-expert bank and lets K3's own routing decide every token, with no experts dropped and no re-quantized substitutes.

The authors frame the work as an experiment rather than a product: if it helps make a frontier model runnable on what they estimate as a $15,000 home setup instead of a $2,000,000 cluster, they consider that worthwhile progress for self-hosted AI. In practice, Deltafin ships as a single native binary built with Rust's cargo toolchain, exposes an OpenAI-compatible server with chat and completion endpoints, and cautions that long conversations run far slower than short completions because prefill and cache costs grow with history.

Why it matters

One token every 3.4 seconds is nowhere near interactive, and the project does not pretend otherwise. The significance is that it runs at all: a checkpoint designed for multi-node GPU clusters now executes end-to-end, unmodified and unquantized, on consumer hardware. That marks out where the practical floor for local inference now sits — fast NVMe, sparse expert routing and speculative decoding together can stand in for a large fraction of the VRAM the model was assumed to need. Publishing the benchmark package and the measurement instruments separately also makes the claims checkable rather than anecdotal. And the underlying problem — deciding which experts to prefetch, where to place them and when to evict them — generalizes to any project trying to run models larger than memory. As storage bandwidth keeps improving, the line between "fits in VRAM" and "runs locally" keeps moving.