// HACKER NEWS — CYBERSECURITY
Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs
TL;DR — Kimi K3 (2.8T-parameter MoE, 1.45 TB of expert weights) running on one M5 Max MacBook Pro with 128 GB, experts streamed from four SSDs.
Measured 2026-09-08 with this fork's configuration of record
(k3-public-bench/env.sh); every number is one cold
run with the exact prompt, and the per-run logs are in
k3-public-bench/results/.
Drive-count ladder on the same prompts: one drive ≈52% of the four-drive
speed, two full mirrors ≈73%, three ≈90% (results/SCALING.md).
Why prefill is slow and what fixes it: results/PREFILL.md.
Definitions, identity scope and precision statement: k3-public-bench/README.md.
Deltafin is a single native binary that runs full Kimi K3. Nothing pruned. Nothing skipped. K3 decides every token.
All 16 experts, every single token. No shortcuts, no "close enough." It's exactly what Moonshot shipped.
The quality rule is simple: K3 itself decides every token, and nobody else. Small draft models are allowed to guess ahead (that's where much of the speed comes from), but K3 checks every guess, and nothing reaches you without its official sign-off.
Pure raw uncut K3 quality, as fast as possible. Speed must never come from reducing model quality. Deltafin keeps all 16 routed experts and the full K3 target as the sole authority for every single token.
Our goal is to squeeze out every last drop of efficiency possible when running a huge model like K3, with all options on the table... except for reducing quality.
Deltafin is not a product pitch. It is an experiment in how far consumer hardware can be pushed, and what we can learn by attempting something so challenging.
Kimi K3 targets infrastructure on the scale of 16 nodes and roughly 4.8 TB of aggregate VRAM. That means the full 2.8T parameters and the 1M-token context window, with the expert bank never pruned. On any home setup, this is an extreme constraint. Every 1% improvement is very hard-won. But each gain can teach something.