According to the source repository, a fork of gavamedia/deltafin enabled the full 2.8‑trillion‑parameter Kimi K3 mixture‑of‑experts model to run on a single M5 Max MacBook Pro equipped with 128 GB of memory. The model’s expert weights, totaling about 1.45 TB, were streamed from four solid‑state drives. Measurements taken on 2026‑09‑08 showed a steady decode rate of 1.00 token/s for a 512‑token answer, 1.13 token/s for 128 tokens, and 0.96 token/s for the public 17‑token prompt (the upstream version had reported 0.68 token/s). When the draft‑assisted decoder was disabled, the rates were 0.9232 token/s for 512 tokens and 0.9261 token/s for 128 tokens; with the draft enabled they rose to 1.0015 token/s and 1.1252 token/s respectively. The time to first token for a 512‑token prompt was approximately 376 seconds (about 6.3 minutes) with the draft off and 375 seconds with it on. The latency was attributed to the prefetch stage re‑reading each layer’s experts eight times; a fix is planned but not yet implemented. Scaling experiments indicated that using one SSD delivered roughly 52 % of the four‑drive speed, two drives about 73 %, and three drives about 90 %. The authors note that the slowest of the 16 parallel expert reads per layer determines the pace, not the total bandwidth. All numbers represent single cold runs with the exact prompts; logs and placement manifests are available in the k3-public‑bench/ directory. The Deltafin binary, distributed under an MIT licence, runs the complete model without pruning or skipping any experts, preserving the exact weights released by Moonshot. The project emphasizes that any speed gains must not come from reducing model quality; draft models may propose tokens ahead of time, but Kimi K3 validates every guess before output. Over recent months the raw throughput improved from 0.0141 token/s on July 27 2026 to 0.2901 token/s on the date of the benchmark, a 1.9 % increase over the previous update. The repository also describes optional components such as the Inferact Kimi‑K3‑DSpark module for chat and a Qwen add‑on that can accelerate raw text continuation by having small models propose continuations that Kimi K3 then checks. Setup instructions include cloning the repository, building the binary with Cargo, and either downloading the full 1.7 TB model to disk or using a streaming mode that fetches experts on demand. An upgrade procedure is provided to preserve existing model data while updating the binary.
