Pipe Network publishes a 350GB Kimi K3 build for 512GB Mac Studios

Why it matters

Pipe made a 1.56TB frontier model loadable on one high-memory Mac, exposing the harder constraint behind open weights: usable local inference still demands severe compromises.

David Rhodus (@DavidRhodus) and Pipe Network have published an Apple MLX port that cuts Moonshot AI's 2.8 trillion-parameter Kimi K3 from roughly 1.56TB on disk to as little as 350GB, allowing a heavily pruned version of the model to load on a Mac Studio with 512GB of unified memory.

Pipe announced the release in a thread on X on July 28th, one day after Moonshot released Kimi K3's weights and technical report. The GitHub repository includes an MLX implementation, a streaming checkpoint converter, calibration tools and scripts for pruning K3's mixture-of-experts architecture. Pipe also uploaded three converted checkpoints to Hugging Face.

The release is a systems engineering project rather than a practical local chatbot. Pipe's fastest published build generates about 0.20 tokens per second, or one token every five seconds. Its larger variants run at 0.14 to 0.16 tokens per second. Pipe describes all three as non-interactive and says pruning causes visible degradation, including repetitive output and responses that drift away from the prompt.

Cutting 896 experts down to 179

Kimi K3 is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters. Moonshot's architecture contains 896 routed experts, selecting 16 for each token, alongside two shared experts. It also combines native image support with a context window of about 1 million tokens, according to Moonshot's technical report.

Pipe's smallest checkpoint keeps 179 of those 896 routed experts, an 80% reduction. The resulting REAP80 build occupies 350GB. A 451GB version retains 242 experts, while a second 451GB checkpoint uses a calibration set weighted toward Chinese and programming code.

The converter processes K3 incrementally because loading the complete checkpoint during conversion would exceed the memory available on a Mac. Pipe says the uncompressed BF16 representation would occupy about 5.6TB. K3's routed experts already use Moonshot's MXFP4 encoding, which MLX can preserve without another lossy quantization pass. Most of the size reduction therefore comes from deleting experts rather than compressing the surviving ones.

That trade-off is substantial. The 451GB model card says the checkpoint loaded on a 512 GiB M3 Ultra in 66 seconds and produced code, English and Chinese samples at roughly 0.16 tokens per second. Pipe warns that the pruning can produce list-like continuations, repetition and weaker general language performance, even when code completion remains structurally coherent.

Pipe's documentation also conflicts with itself. The individual Hugging Face cards report successful generation and measured speeds on a 512 GiB M3 Ultra. A separate "Reality check" section in the GitHub README says no published tier runs on a single Apple Silicon machine and that the artifacts have never produced a token. Elsewhere in the same README, Pipe again lists the 350GB and 451GB builds with measured throughput. The Hugging Face cards contain the more specific test configuration and output samples, but the contradiction makes independent reproduction essential.

Rhodus turns a CDN company toward AI infrastructure

Rhodus built Pipe through Permissionless Labs, which now describes itself as an AI infrastructure lab offering inference, storage and fine-tuning. That positioning is a marked expansion from Pipe's original pitch as a decentralized content delivery network using independently operated nodes to move data closer to users.

Rhodus spent roughly two decades in streaming and data infrastructure before starting Pipe. He was CTO of Volar Video and an early employee at Elemental Technologies, the video infrastructure provider acquired by Amazon Web Services. He later worked at ConsenSys, according to Multicoin Capital's account of its investment.

In September 2024, Permissionless Labs raised a $10 million Series A led by Multicoin, with Robot Ventures, Solana Ventures, Solana co-founder Anatoly Yakovenko and Meltem Demirors participating. The financing was presented at the time as capital for expanding Pipe's CDN network, recruiting engineers and launching its testnet.

The Kimi port shows Rhodus applying the same underlying bet to AI: large models are becoming downloadable, while moving, storing and serving their weights remains an infrastructure problem. Shrinking K3 to fit inside Apple's largest unified-memory configuration crosses a technical threshold. At 0.20 tokens per second and with most routed experts removed, it remains an experiment for developers willing to trade speed and model quality for control over the machine running it.