Skip to content
modelread

Loading devices…

Remembered on this device. Multi-GPU: search “2x”, “8x”, “node”.

HTTP API

Base URL https://modelread.com. JSON in and out, no key. Requests are rate-limited per client (about 20 per minute); a 429 means wait a minute. The machine-readable schema is /openapi.json.

GET /api/estimate

ParameterType / defaultMeaning
repostring, requiredHugging Face repo id
hwstring · rtx3090Device (see Hardware strings)
ctxint · 8192Context length
ramfloatSystem RAM / unified memory override, GiB
kvf16 | q8_0 | q4_0 · f16KV cache type
visionbool · falseInclude the vision tower

200 returns an estimate (see Reading results). 202 means the repo is new to the service and is being read from the Hub; retry after about a minute.

curl "https://modelread.com/api/estimate?repo=openai/gpt-oss-120b&hw=dgx-spark&ctx=65536&kv=q8_0"

POST /api/ask/stream

Body {"q": "...", "hw": "<preset id, optional>", "ram": <GiB, optional>}. Server-sent events:status (progress text), query (how the question was understood), draft (markdown),final (answer_md, query, notes, results[], queued[]) or error. hw is used only when the question names no device.

curl -N "https://modelread.com/api/ask/stream" -H 'content-type: application/json' \
  -d '{"q":"best coding model","hw":"m4max-128"}'

POST /api/ask

Same as the stream, as one JSON response. Add "llm": false for the deterministic ranking only.

GET /api/hardware

Every device preset: id, name, kind, segment, vendor, VRAM, RAM, bandwidth, FP16 TFLOPS, backend, GPU count, $/hour for cloud instances.

GET /api/matrix?hw={id}

The nightly “what runs where” table: best variant per popular model on reference devices at 32K context.

GET /api/search

ParameterType / defaultMeaning
taskstring · chatTask family
qstringName text
matchbool · falseRequire every word of q in the repo id
max_params_bfloatSize bound
limitint · 20≤ 100

GET /api/facts?repo=…

The raw facts bundle (files, architecture, quantized derivatives) that the in-browser WASM estimator consumes.

GET /api/changes?days=14

New and changed models in the index.

POST /api/measurements

Submit a measured run (usually via modelfit bench --submit). Validated and screened for outliers before it is used for fitting.

Askwhat should I run?
Browsebest models per device
Devicesall hardware
Connect your agentMCP · API
How the numbers workcalibration