← Back to articles

Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs

2026-09-08 | original

read original get Samsung T7 Shield Portable SSD → more articles

Why This Matters

This project shows a 2.8-trillion-parameter frontier-scale model running unpruned on a consumer MacBook Pro by streaming weights from four SSDs, hitting roughly 1 token/s. It matters because it reframes the hardware floor for running the largest open-weight models: storage bandwidth and clever scheduling, not just VRAM, become the bottleneck. Even at unusable speeds for chat, it's a useful benchmark of how far local inference engineering can be pushed without quality loss.

Key Takeaways

Deltafin runs full Kimi K3 with all 16 routed experts intact, using speculative draft models only where K3 verifies every token, so no quality is traded for speed.
Throughput on an M1 Max has improved dramatically over successive updates, from 0.0141 to ~0.29 tokens/s in about a week of optimization work.
The approach makes SSD streaming and weight placement a first-class inference problem, and the fork publishes benchmarks, placement manifests and measurement tooling openly.

Worth a Look

Samsung T7 Shield Portable SSD — Streaming a 2.8-trillion-parameter model off disk lives or dies on SSD throughput, and the T7 Shield is a rugged, bus-powered USB-C drive that plugs straight into a MacBook Pro. Grab several and you can mirror the multi-drive setup the Deltafin benchmark uses — or just enjoy fast external storage for huge model weights.

See Samsung T7 Shield Portable SSD on Amazon → Affiliate link — we may earn a commission on purchases, at no extra cost to you. Product picked by AI based on this article; it is not a tested recommendation.

ARGODRIVE Deltafin

A fork of gavamedia/deltafin (MIT) running Kimi K3 from SSDs on Apple Silicon, with the ARGODRIVE storage work. The benchmark package, placement manifests and results are in k3-public-bench/ ; the measurement instruments are published separately as ARGODRIVE. Credits and what this fork changes: CREDITS.md . The upstream README follows.

____ _ _ __ _ | _ \ ___| | |_ __ _ / _(_)_ __ | | | |/ _ \ | __/ _` | |_| | '_ \ | |_| | __/ | || (_| | _| | | | | |____/ \___|_|\__\__,_|_| |_|_| |_| Run the full, never-pruned, 2.8-trillion-parameter Kimi K3 on consumer hardware, as "fast" as possible

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.

Latest Benchmarks on an M1 Max laptop

0.2901 token/s (3.447 s/token) — 1.9% higher throughput than last update

Historical M1 benchmarks:

0.2847 token/s (August 2, 2026) — 7.0% higher throughput

... continue reading

Explore topics: kimi k3 deltafin argodrive apple silicon m1 max