{
  "claim": "Edge0 publishes five different decode rates and four different peak-memory figures for the same 35B tier, and two of the decode figures claim the same machine. Nothing in the release says which supersedes which. The spread on decode is 13.0 to 20.4 tok/s — 57% — and the README's headline prerouter gain (\"up to +59%\") appears nowhere in the paper it links to.",
  "method": "Read out of the four places Edge0 publishes numbers, all at the same commit (5b65e6f, 18 Sept 2026): the GitHub README, the Hugging Face model card, arXiv:2609.18063v2 (paper/main.pdf in the repo, identical to the arXiv listing), and the pinned acceptance constants in src/edge0/models/edge0_35b/__init__.py. The iPhone figures are read off the on-screen overlay of 20260910-105854.mp4, the 89-second screen recording committed to the Hugging Face repo, at 0:12.",
  "source": "https://arxiv.org/abs/2609.18063",
  "captured": "2026-09-18",
  "columns": [
    { "key": "quantity", "label": "quantity" },
    { "key": "value", "label": "published value", "align": "right" },
    { "key": "machine", "label": "machine" },
    { "key": "where", "label": "where it is published" }
  ],
  "rows": [
    { "quantity": "edge0-35b decode", "value": "13.0 tok/s", "machine": "unstated", "where": "src/edge0/models/edge0_35b/__init__.py — target_tok_s=13.0" },
    { "quantity": "edge0-35b decode", "value": "14.9–17.7 tok/s", "machine": "Mac mini M4 Pro, 24 GB", "where": "README.md Benchmark table; HF card Performance table" },
    { "quantity": "edge0-35b decode", "value": "15 tok/s", "machine": "unstated", "where": "HF card headline + \"Fast enough for interactive use\"" },
    { "quantity": "edge0-35b decode", "value": "~16 tok/s", "machine": "iPhone 16 Pro, 8 GB", "where": "20260910-105854.mp4 overlay, 0:12 (HF repo)" },
    { "quantity": "edge0-35b decode", "value": "20.4 tok/s", "machine": "Mac mini M4 Pro, 24 GB", "where": "arXiv 2609.18063v2 Table 1; abstract says \"≈20 tok/s\"" },
    { "quantity": "edge0-35b peak memory", "value": "1.7 GiB / 1.8 GB", "machine": "iPhone 16 Pro, 8 GB", "where": "20260910-105854.mp4 overlay + caption" },
    { "quantity": "edge0-35b peak memory", "value": "2.9 GiB", "machine": "Mac mini M4 Pro, 24 GB", "where": "README.md; HF card; arXiv Table 1 — all agree" },
    { "quantity": "edge0-35b peak memory", "value": "3400 MB (3.32 GiB)", "machine": "unstated", "where": "src/edge0/models/edge0_35b/__init__.py — peak_active_mem_mb=3400.0" },
    { "quantity": "edge0-8b decode", "value": "23.9–25.3 tok/s", "machine": "Mac mini M4 Pro, 24 GB", "where": "README.md Benchmark table; HF card" },
    { "quantity": "edge0-8b decode", "value": "25.0 tok/s (40.02 ms/step)", "machine": "Mac mini M4 Pro, 24 GB", "where": "src/edge0/streaming/options.py — prod_k8 docstring" },
    { "quantity": "edge0-8b decode", "value": "28.0 tok/s", "machine": "Mac mini M4 Pro, 24 GB", "where": "arXiv 2609.18063v2 Table 1" },
    { "quantity": "edge0-8b decode", "value": "33.0 tok/s", "machine": "unstated", "where": "src/edge0/models/edge0_8b/__init__.py — target_tok_s=33.0" },
    { "quantity": "edge0-8b peak memory", "value": "1.0 GiB", "machine": "Mac mini M4 Pro, 24 GB", "where": "README.md; HF card headline \"1 GiB active memory\"" },
    { "quantity": "edge0-8b peak memory", "value": "1.4 GiB", "machine": "Mac mini M4 Pro, 24 GB", "where": "src/edge0/streaming/options.py — prod_k8 docstring" },
    { "quantity": "edge0-8b peak memory", "value": "1400 MB (1.37 GiB)", "machine": "unstated", "where": "src/edge0/models/edge0_8b/__init__.py — peak_active_mem_mb=1400.0" },
    { "quantity": "edge0-8b peak memory", "value": "1.5 GiB", "machine": "Mac mini M4 Pro, 24 GB", "where": "arXiv 2609.18063v2 Table 1" },
    { "quantity": "prerouter decode gain", "value": "+2.5% (19.9 → 20.4 tok/s)", "machine": "Mac mini M4 Pro, 24 GB", "where": "arXiv 2609.18063v2, Table 1 caption" },
    { "quantity": "prerouter decode gain", "value": "~5%", "machine": "Mac mini M4 Pro (8B tier)", "where": "src/edge0/streaming/options.py — prod_k8 docstring" },
    { "quantity": "prerouter decode gain", "value": "up to +59%", "machine": "unstated", "where": "README.md and HF card — not reproduced in the paper" },
    { "quantity": "prerouter decode gain", "value": "+80 / +82 / +84% (K=2/4/8)", "machine": "MacBook M2, 16 GB (checkpoint does not fit)", "where": "arXiv 2609.18063v2 Table 3 and Figure 3a" }
  ],
  "note": "The 35B tier's routed width is the other thing worth reading twice. The checkpoint's own config.json sets num_experts_per_tok: 8; edge0 serves it with MoESpec(top_k=4). Measured from the tensor headers, the active set is 2,443,115,136 parameters at K=4 against 2,946,431,616 at K=8 — so the \"A3B\" in Edge0-35B-A3B-preview describes the base model's native width, not the one that ships. The 8B tier's \"A1B\" does check out: 1,170,908,064 active at its native K=8, and it only reaches 1.2B because tie_word_embeddings is false and the 241,434,624-parameter lm_head really is a separate tensor in the file."
}