The Data API
Stable, versioned JSON for the community dataset โ no server, no auth
The 30-second version: Everything this site shows is public JSON. Fetch data.json from any script or website, check its version field, validate it against the JSON Schema, and you're building on the community's real hardware/model/performance data. No server, no auth, no key โ just fetch().
What this is
The site is fully static โ there is no backend. The "API" is the dataset itself, published at stable URLs and regenerated by the automated weekly pull:
data.jsonโ the canonical submissions dataset (what the site's own tools use)data/api/v1/submissions.jsonโ the same payload at its versioned pathdata/api/v1/schema.jsonโ the JSON Schema (draft 2020-12) that formally defines the shapedata/api/v1/benchmarks.json/data/api/v1/hardware.jsonโ the secondary community datasets (structured since v1.4)data/api/v1/benchmarks_schema.json/data/api/v1/hardware_schema.jsonโ their JSON Schemas
The full contract โ every field, the versioning policy, the changelog โ is documented in the API README in the repo.
The shape of a record
Each submission in submissions[] carries the 20-key canonical shape โ provenance, hardware, OS/driver, model, quant, context, performance, and curation metadata:
{
"version": "1.4",
"generated_at": "2026-10-26T12:13:38.524400",
"schema_version": "2.4",
"source_thread": "https://forum.level1techs.com/t/256443",
"total_submissions": 50,
"skipped_templates": 3,
"submissions": [
{
"id": "post_2_madgoat_1",
"source": { "post_number": 2, "post_id": 4119054, "username": "MadGoat" },
"type": "full",
"tier": "t1",
"hardware": { "description": "1x Radeon VII 16GB HBM2 โฆ", "vendor": "AMD" },
"backend": "llama-swap (llama.cpp subprocess, ROCm)",
"os_driver": "Linux/ROCm",
"model": { "name": "samuelcardillo/Carnice-Qwen3.6-MoE-35B-A3B" },
"quant": "GGUF I-Nano",
"context": { "stable": "65536 tokens", "kv_cache": "Q8_0 (quantized)" },
"performance": { "decode_tps": "54.4 decode", "prefill_tps": "215" }
}
]
}
(abbreviated โ the schema defines every field, type, and closed enum precisely.)
The secondary datasets
Two more community threads are published the same way, now as structured data (v1.4):
data/api/v1/benchmarks.jsonโ the Crucible benchmarking thread: each item carriesmodel,hardware,backend,performance(decode / prefill / VRAM / TTFT),context_length, andpower_wโ allnullwhen the post doesn't state them.data/api/v1/hardware.jsonโ the "Show Off Your AI Rig" thread: each item carriesgpu(model + count),vram_total_gb,cpu,ram_gb,storage,os, andworkload(what it runs, with verbatim speed snippets).
The extraction is deliberately conservative: quoted content is excluded (a reply quoting someone else's rig doesn't credit the replier), questions and hypotheticals aren't specs, and nothing is guessed โ a field that isn't in the post stays null. Every item keeps its 300-char preview as the fallback, and a structured flag tells you which items carry real data. Both payloads validate against their schemas (benchmarks_schema.json / hardware_schema.json), and each structured value is verbatim from the source post โ the notes field carries the sentence it came from.
Versioning
The version field is a major.minor stamp. Additive changes (new optional fields, new enum values) bump the minor in place; breaking changes bump the major and publish under a new data/api/vN/ path โ old versions stay frozen, so your code never breaks silently. Pin the major and check version at runtime.
Use it from your own site
GitHub Pages serves every file with Access-Control-Allow-Origin: *, so any page on any domain can fetch the data straight from the browser:
const res = await fetch("https://madgoathaz.github.io/llm-matrix/data.json");
const data = await res.json();
console.log(data.version, data.submissions.length);
The data changes at most weekly (the automated pull runs Mondays), so cache it.
The honest part
The performance and context fields are community-reported strings, not lab measurements โ "54.4 decode" is what a person typed, and that's the point of the project: real people, real hardware, real numbers. Parse the leading number when you need one; treat confidence as your quality signal.
Next steps
Read the full contract. Every field, the versioning policy, and the changelog live in the API README in the repo.
Validate against the schema. The formal JSON Schema (draft 2020-12) is served at data/api/v1/schema.json โ point your validator at it.
Fetch the data. data.json (site root) or data/api/v1/submissions.json โ same payload, the versioned path is the one to code against.