Kimi K3 on a MacBook Pro: SSD Cost Reality

The independent Deltafin project has demonstrated Kimi K3 running on a MacBook Pro by streaming its enormous expert bank from four SSDs. Its headline is real but narrow: 1.0015 tokens per second of steady decode over one 512-token raw completion on an Apple M5 Max with 128GB of unified memory.

This is an engineering experiment, not a new Moonshot AI model, endpoint, or price. The project also reports roughly 375 seconds to the first token for a 512-token prompt. Buyers should read the result as proof that consumer-class Apple Silicon can execute the model—not as evidence that a laptop replaces a production API.

What the 1 token/s result actually measured

Deltafin published its configuration, benchmark scripts, placement manifests, per-run logs, and measurement definitions. The September 8 result used the internal SSD plus three NVMe enclosures, with a complete expert copy on one 2TB external drive and usage-weighted replicas across the others.

Deltafin M5 Max test	Project-reported result
Steady decode, 512 generated tokens, drafter on	1.0015 tok/s
Inclusive throughput, 512 generated tokens	0.9849 tok/s
Steady decode, 128 generated tokens	1.1252 tok/s
First token, 512-token prompt	About 375 seconds
Four-drive expert reads	About 24 GB/s average

“Steady” excludes first-token processing. “Inclusive” starts at process launch and includes the cold load. The 1 tok/s headline therefore describes sustained generation after a long startup, not the wait a user feels before an answer begins.

The project reports medians of two cold runs per setting and notes modest variation after large staging writes. It remains one machine and a small prompt set.

Four SSDs materially change throughput

On the 128-token drafted test, throughput rose from 0.5350 inclusive tok/s with one drive to 1.0377 with four. The one- and two-drive tests used full mirrors; three and four used placement based on recorded expert usage.

The MacBook held a 50.7GiB int8 resident spine in unified memory. The routed Kimi K3 expert files stayed at the released MXFP4 precision, but the project explicitly does not claim BF16 bit-exactness for the full execution path. A Qwen3-0.6B drafter proposed tokens, and Kimi K3 checked them before acceptance.

That distinction matters because Kimi K3 is a sparse mixture-of-experts model. Moonshot’s official model card lists 2.8 trillion total parameters, 104 billion activated parameters, 896 routed experts, and 16 selected experts per token. Deltafin keeps the expert bank available on storage; it does not make all 2.8T parameters resident in the Mac’s unified memory.

What the million-token pace means

At the measured 1.0015 tok/s steady rate, generating one million output tokens would take about 277 hours, or 11.6 days, before accounting for repeated first-token waits, longer prompts, failures, maintenance, or idle time. At uninterrupted steady decode, the machine would produce roughly 31.6 million output tokens per year.

The live chart above prices managed traffic from the maintained dataset rather than freezing an API rate in this article. That still is not a complete self-hosting comparison: the local experiment trades token billing for a MacBook, multiple SSDs and enclosures, electricity, setup time, operational work, and very low interactive throughput.

The project itself frames Deltafin as research. Its home-setup estimate is an ambition, not an audited bill of materials for this exact benchmark, so it cannot establish a break-even point.

Did Kimi K3 API pricing change?

No. Moonshot’s official rate card still lists Kimi K3 per million tokens with separate cache-hit input, cache-miss input, and output prices. The live cards above are rendered from our daily-maintained pricing dataset, and the machine-readable API data contains the same rows.

Telnyx and Novita are separate managed routes with their own serving terms. Teams that want Kimi K3 without reproducing a four-drive local setup can compare Novita’s managed Kimi K3 route with Moonshot and Telnyx.

Affiliate disclosure: we may earn a commission from the sponsored link above. It does not affect the benchmark or pricing analysis.

Labs inclusion and the local-hardware blocker

Kimi K3 is already included in AI Pricing Guru Labs. Its complete managed-route run attempted 49 deterministic text tasks, returned no endpoint errors, and answered 47 correctly. That result measures cost per correct answer through a public endpoint.

The Deltafin result remains a separate local-throughput coverage note. A fair Labs comparison would need the exact repository commit and weight artifacts pinned, repeat runs on independently controlled hardware, identical prompts and output lengths across local and managed routes, warm and cold latency, energy measurements, failure rates, and a full hardware-cost ledger. The current text leaderboard cannot turn a storage-specific tok/s result into an API score.

Buyer takeaway

Deltafin makes an important technical point: fast local storage and MoE-aware scheduling can let a 2.8T open-weight model run on a 128GB Mac. It also makes the commercial limitation visible. Six-minute first-token latency and around one output token per second fit experimentation, offline research, and storage-engine work—not interactive production serving.

Use the Moonshot AI pricing page and token cost calculator for managed K3 workloads. Compare smaller hosted open models on the Together AI pricing page, and use our local AI versus API guide to structure the full break-even decision. Consider the local route when privacy, offline access, systems research, or control matters enough to justify the hardware and engineering effort.

Sources: the independent ARGODRIVE Deltafin repository, its published benchmark package, the Hacker News discussion, Moonshot AI’s official Kimi K3 model card, and the official Kimi K3 API pricing page. Facts checked September 8, 2026 at 22:18 UTC.