MiMo-V2.6 Flash

Price

from catalog

Context and output

Context window262,144Tokens per request

Max output131,072Tokens per response

Everything you send counts toward the context window — system prompt, conversation history, retrieved context and tool definitions. Both figures are read from the catalog, so they cannot drift from what the gateway enforces.

Rate limits and quotas

Anonymous playground8 / s burst · 60 / min · 300 / hour · 1,000 / dayApplies only to anonymous requests without an API key. Registered free accounts have their own policy below.

Funded accountsAccount limitsPaid requests use your account’s limits, visible in the console.

Non-streaming timeout~100 s first byteNon-streaming requests pass through a CDN. Long outputs must use stream=true.

Guest playgroundNo key requiredThe box below calls the same gateway under the free-use caps. No sign-up, no card.

Concurrency64 / key, defaultEnforced per API key as a concurrency lease; the default is 64. Account and per-key limits can be adjusted in the console. Free and keyless requests are not governed by this row.

Benchmarks

Being measured

First-token latency—Will state the precision served, concurrency, prompt shape, timestamp

Output throughput—Tokens per second, single stream

90-day uptime—Per service, planned windows marked separately

Load reportsWhat a published number will carry

Current live status: Status page

Try it

Ask anything. The response streams token by token from the same gateway your code will call.

Copy setup for your AI assistant

Installs the SDK, sets base_url and the model id, reads the key from an environment variable, writes a streaming smoke test.

Quickstart

Data and hardware

Data handlingNothing is written to disk

Inference runs entirely in memory and is erased on completion. Only request metadata — token counts, timestamps, latency, routing — is retained, for up to 90 days, for billing and reliability.

Transparency hub

HardwareHardware we operate

This model is served at FP8 on hardware we operate, with no silent downgrade under load. Execution happens in one published region, the United States, and the region each request ran in is recorded on that request.