How fast does full Kimi K3 2.8T run on a MacBook?

Benchmarks published September 8, 2026 measured the full unpruned 2.8T-parameter Kimi K3 at 1.0015 tokens/s on an M5 Max with 128 GB of RAM, streaming experts from four SSDs — 0.9232 tok/s without speculative drafting, up from 0.0141 tok/s six weeks earlier.

Why stream model experts from SSDs instead of quantizing them?

Quantizing to fit degrades the weights and the cost is rarely measured. A Qwen3.8 27B benchmark published the same week found 4-bit quantization holds up but 1-bit collapses — capability falls off a cliff at aggressive bit rates. Streaming the untouched weights from SSDs preserves exact model quality at the cost of speed and drives.

How does SSD count affect local LLM decode speed?

The Deltafin drive ladder measured one SSD at roughly 52% of four-drive throughput, two mirrored pairs at 73%, and three drives at 90%. For MoE models too large for RAM, decode speed scales nearly linearly with aggregate SSD bandwidth.

Kimi K3 2.8T Runs at 1 Token/s on a MacBook From Four SSDs

Deltafin hit 1.0015 tokens/s running the full, never-pruned 2.8-trillion-parameter Kimi K3 on an M5 Max with 128 GB — every expert byte exactly as Moonshot shipped, streamed from four SSDs. Benchmarks published today. Six weeks ago this setup ran 0.0141 tok/s. That's a 70x climb on unmodified open weights.

The connection nobody has made yet: this is the same week Quesma benchmarked Qwen3.8 27B quantizations and found 4-bit holds up but 1-bit collapses. Two camps, same problem — frontier open weights that don't fit your hardware — and only one of them is safe.

What shipped

The ARGODRIVE fork's measurement package is unusually honest: one cold run per prompt, per-run logs published, definitions stated up front. Drafter off: 0.9232 tok/s over 512 tokens. Drafter on: 1.0015 — the small model guesses, but K3 itself verifies every token, so no quality leaves the room. TTFT is ≈375s; prefill is the remaining wall. The most useful number is the drive-count ladder: one SSD ≈ 52% of four-drive throughput, two mirrored pairs ≈ 73%, three ≈ 90%. Decode speed here is an I/O problem wearing a model costume — you're buying tokens with SSD bandwidth almost linearly.

Why it matters: stream vs. shrink

Deltafin's README takes a direct shot at competing setups that re-encode K3's expert bank to ~3 bits — and notes nobody has measured what that costs. Quesma's data today is the first real receipt for that suspicion: at aggressive quantization, capability doesn't degrade gracefully, it falls off a cliff. The stream-untouched path sidesteps the question entirely: weights stay exactly as released, and the price is paid in hardware ($500 of NVMe instead of nothing) and patience (one token per second, 6-minute time-to-first-token).

Verdict: 1 tok/s of full-precision 2.8T beats 5 tok/s of a model you can't vouch for. The open-weights middle ground now has two roads — compress and hope, or stream and wait — and today's evidence favors the second for anything where correctness matters. Kimi's own deployment target is 16 nodes and ~4.8 TB of VRAM; a laptop and four drives just closed a four-hundred-to-one gap.

What shipped

Why it matters: stream vs. shrink

Related