Skip to main content
This benchmark measures how quickly streaming begins on the public MKA1 Responses API. It uses POST /api/v1/llm/responses on https://apigw.mka1.com/api/v1/llm/responses with stream: true and model meetkai:functionary-pt. The goal is simple: measure the time from sending the request until the first streamed text token reaches the client.

​
Results summary

Benchmark run: 2026-03-31 On this run, the public Responses API stayed below 1 second at the 95th percentile for time to first streamed text token.

​
What TTFT means in this benchmark

For this benchmark, TTFT means:
  • the client starts a timer immediately before the HTTP request is sent
  • the client opens the SSE stream
  • the timer stops when the first non-empty streamed text delta is received
This model typically emits reasoning text before the final answer text. Because of that, the benchmark counts the first non-empty text delta from either:
  • response.reasoning_text.delta
  • response.output_text.delta
That definition measures the start of continuous streamed text as seen by the client. The benchmark also reports two earlier milestones:
  • response headers: when the initial HTTP response is returned
  • first SSE event: when the first SSE frame arrives, regardless of content

​
Benchmark setup

The live run used the following request profile:
  • base URL: https://apigw.mka1.com
  • endpoint: POST /api/v1/llm/responses
  • auth: Authorization: Bearer <mka1-api-key>
  • end-user header: X-On-Behalf-Of: ttft-benchmark-user
  • model: meetkai:functionary-pt
  • prompt: Reply with exactly the text OK and nothing else.
  • stream: true
  • temperature: 0
  • max_output_tokens: 64
  • warmups: 5
  • measured requests: 100
  • concurrency: 1
  • request timeout: 45s
These numbers are client-observed measurements against the public API endpoint. They therefore include network round-trip time between the benchmark client and apigw.mka1.com, not just server-side processing time.

​
Request shape

The benchmark itself uses an SSE parser rather than curl output timing so it can detect the exact arrival time of the first streamed text token.

​
Interpretation

This run shows three useful layers of latency:
  • the connection and gateway returned headers quickly, with p95 at 70.42 ms
  • the first SSE frame arrived at p95 70.69 ms
  • the first streamed text token arrived at p95 608.49 ms
That means the stream starts almost immediately at the transport level, and meaningful streamed text reaches the client well under 1 second at p95 for this model and request profile.

​
Full Test output