Argonaut Labs · 2026-09-08 · notable
Deltafin — Kimi K3's 2.8T weights stream off four SSDs at 1 token/s
Deltafin runs the full 2.8-trillion-parameter Kimi K3 on one MacBook Pro. Argonaut Labs' ARGODRIVE build streams 1.45 TB of expert weights from four SSDs into 128 GB of RAM and decodes about 1 token per second.
A Rust binary that streams a 2.8T-parameter MoE off SSDs so it runs on a laptop.
Key specs
| Decode speed | 1.00 tok/s |
|---|---|
| Weights streamed | 1.45 TB |
What is it?
Deltafin runs the complete, unmodified Kimi K3 — 2.8 trillion parameters — on a single M5 Max MacBook Pro with 128 GB of RAM. Argonaut Labs' ARGODRIVE build adds the storage work and a benchmark package on top of gavamedia's original Deltafin. All 16 experts still process every token, so the model is not pruned or distilled down to fit the machine.
How does it work?
Streaming replaces loading. The weights sit on four SSDs and each expert is pulled in on demand, with router tracing and expert prefetch used to hide the read latency. The single compiled Rust binary also handles native tokenization, and an optional smaller draft model called DSpark speeds up decoding, with K3 verifying every token DSpark proposes.
Why does it matter?
Throughput lands at about 1 token per second — 1.13 tok/s on a 128-token answer and 1.00 tok/s on a 512-token one — which is slow for chat but shows a frontier-scale MoE running on hardware a person can own. Drive count matters: one SSD gives roughly 52% of the four-drive speed, two give 73% and three give 90%. Time to first token is still about 6.3 minutes on a 512-token prompt, which the README lists as an open issue.
Who is it for?
local-LLM tinkerers on Apple Silicon
Try it
cargo build --locked --release && ./target/release/deltafin run --chat