mirror of
https://github.com/hanzo-apps/hanzo.ai.git
synced 2026-10-28 07:32:52 +00:00
zeekayandHanzo Dev
dd3507e28b
The index carried a frontier grid ranked first on every benchmark, 1,800
tok/s, a Metal ratio of 1.024 against 0.967 in the data, a LongMemEval
baseline of 91.8 against 93.4, "125 ns sandboxes" for a goroutine spawn, an
adversarial 0.706 and "zero model invocations". None had a run behind it.
The index now reads INDEX in lib/bench-data, which resolves each result by
name from lib/data: MemoryAgentBench pooled and 262k, LoCoMo ALL@20 and R@20,
LongMemEval ALL@5, RepoBench-R R@1, CronQuestions latency at equal Hits@1,
and the runtime board, now lib/data/benchmarks-runtime.json. GPQA is not
quoted there; its page states that the Enso paper's own run disagrees.
The loss sections d1d6de683 removed are back on every page. The dead
index components, lib/runtime.ts and the unused platform helpers are gone.
e2e/unit/figures.spec.ts fails on any digit in the index source, and
e2e/benchmarks.spec.ts checks the render against INDEX, the withdrawn
claims, and each page's loss heading.
Co-authored-by: Hanzo Dev <[email protected]>
135 lines
3.0 KiB
JSON
135 lines
3.0 KiB
JSON
{
|
||
"passes": {
|
||
"fleet": {
|
||
"repo": "hanzoai/benchmarks",
|
||
"url": "https://github.com/hanzoai/benchmarks",
|
||
"board": "README.md",
|
||
"when": "between 2026-09-07 and 2026-09-11",
|
||
"machine": "an M-series laptop",
|
||
"command": "node fleet/fleet.mjs /tmp/fleet"
|
||
},
|
||
"goroutine": {
|
||
"repo": "hanzoai/benchmarks",
|
||
"url": "https://github.com/hanzoai/benchmarks",
|
||
"board": "README.md",
|
||
"when": "between 2026-09-07 and 2026-09-11",
|
||
"machine": "an Apple M4 Max, 16 cores",
|
||
"command": "cd goroutine && go build -o /tmp/g . && /tmp/g"
|
||
},
|
||
"sandbox": {
|
||
"repo": "hanzoai/benchmarks",
|
||
"url": "https://github.com/hanzoai/benchmarks",
|
||
"board": "README.md",
|
||
"when": "between 2026-09-07 and 2026-09-11",
|
||
"machine": "an M-series laptop",
|
||
"command": "node sandbox/sandbox.mjs"
|
||
}
|
||
},
|
||
"shots": [
|
||
{
|
||
"part": "fleet",
|
||
"what": "state per dormant agent",
|
||
"ours": "477 bytes",
|
||
"theirs": "~1 MB, a competitor’s own modelled assumption",
|
||
"checked": null
|
||
},
|
||
{
|
||
"part": "fleet",
|
||
"what": "one million dormant agents on disk",
|
||
"ours": "455 MB",
|
||
"theirs": "~977 GB at that assumption",
|
||
"checked": null
|
||
},
|
||
{
|
||
"part": "fleet",
|
||
"what": "dormant agents written per second",
|
||
"ours": "309,789",
|
||
"theirs": null,
|
||
"checked": null
|
||
},
|
||
{
|
||
"part": "fleet",
|
||
"what": "resume one, in process",
|
||
"ours": "0.034 ms",
|
||
"theirs": null,
|
||
"checked": null
|
||
},
|
||
{
|
||
"part": "goroutine",
|
||
"what": "heap per live agent",
|
||
"ours": "601 bytes",
|
||
"theirs": "1.2 MB per isolate, isolated-vm",
|
||
"checked": "1.00 MiB"
|
||
},
|
||
{
|
||
"part": "goroutine",
|
||
"what": "one million live agents",
|
||
"ours": "573.0 MB",
|
||
"theirs": null,
|
||
"checked": null
|
||
},
|
||
{
|
||
"part": "goroutine",
|
||
"what": "spawn one",
|
||
"ours": "125 ns",
|
||
"theirs": "2.79 ms, isolated-vm",
|
||
"checked": "0.35 ms"
|
||
},
|
||
{
|
||
"part": "goroutine",
|
||
"what": "wake one million",
|
||
"ours": "107 ms — 107 ns each",
|
||
"theirs": null,
|
||
"checked": null
|
||
},
|
||
{
|
||
"part": "sandbox",
|
||
"what": "V8 context",
|
||
"ours": "0.13 ms",
|
||
"theirs": null,
|
||
"checked": null
|
||
},
|
||
{
|
||
"part": "sandbox",
|
||
"what": "V8 isolate, isolated-vm 7.0.1",
|
||
"ours": null,
|
||
"theirs": "2.79 ms, 1.2 MB",
|
||
"checked": "0.35 ms, 1.00 MiB"
|
||
},
|
||
{
|
||
"part": "sandbox",
|
||
"what": "hanzo-vm, cold boot",
|
||
"ours": "309 ms",
|
||
"theirs": "E2B under 200 ms · Modal ~1 s · Cloudflare 1–3 s",
|
||
"checked": null
|
||
},
|
||
{
|
||
"part": "sandbox",
|
||
"what": "hanzo-vm, from a checkpoint",
|
||
"ours": "311 ms",
|
||
"theirs": "Morph under 250 ms resume · E2B ~1 s from pause",
|
||
"checked": null
|
||
}
|
||
],
|
||
"hosts": [
|
||
{
|
||
"runtime": "wazero",
|
||
"language": "WASM — any language that targets it",
|
||
"instantiate": "7.8 µs",
|
||
"call": "20 ns"
|
||
},
|
||
{
|
||
"runtime": "goja",
|
||
"language": "JavaScript",
|
||
"instantiate": "2.4 µs per VM",
|
||
"call": "740 ns warm eval"
|
||
},
|
||
{
|
||
"runtime": "gpython",
|
||
"language": "Python",
|
||
"instantiate": "25.0 µs per context",
|
||
"call": "—"
|
||
}
|
||
]
|
||
}
|