MiniMax M3 Free on OpenRouter: We Gave It a Real Coding Task

MiniMax passed this one-shot Next.js task in 4m 46.944s: the production build passed, and all four required issue-inbox interactions worked.

Desktop screenshot of the responsive Issue Inbox MiniMax built, showing the search field, Clear button, live six-of-six count, and six seeded issue rows. — MiniMax M3 FreeIssue Inbox, desktopOpen full capture

Mobile screenshot of the Issue Inbox MiniMax built, with the search controls stacked above six responsive issue rows. — MiniMax M3 FreeIssue Inbox, mobileOpen full capture

What happened

MiniMax completed the bounded contestant run successfully. Hermes exited with code 0 after 286,944 milliseconds (4 minutes 46.944 seconds), without reaching the 10-minute operator wall.

The untouched submission passed its production build. Browser validation found six issues initially, filtered the list to the matching dark-mode issue, updated the count from “Showing 6 of 6 issues” to “Showing 1 of 6 issues,” and restored all six issues when Clear was pressed.

The same frozen submission rendered at desktop and mobile sizes with no browser console errors and no material mobile horizontal overflow.

Hermes recorded 36 successful MiniMax model calls, zero HTTP 429 responses, and zero seconds of provider backoff. No run-level monetary cost was exposed, so none is estimated.

Meaningful differences

Difference	Observed
Requested functionality	All four required interaction checks passed on the untouched submission: six initial issues, title search filtering, a live visible-count update, and Clear restoring the full list.
Responsive output	The preserved desktop and mobile captures show the same Issue Inbox adapted to both viewport sizes; the mobile check measured zero pixels of horizontal overflow.
Execution reliability	This clean run finished in one contestant invocation with exit code 0, no HTTP 429 responses, no provider backoff, no contestant repair pass, and no post-submission source modification.

What this tells us

In this single bounded test, MiniMax M3 Free through OpenRouter, operating through Hermes on a VPS, successfully completed this specific small Next.js coding task under the stated constraints.

Evidence and setup

Setup

{
  "participants": [
    "MiniMax M3 Free"
  ],
  "model": "minimax/minimax-m3:free",
  "provider": "OpenRouter",
  "agent_harness": "Hermes",
  "execution_environment": "VPS",
  "held_constant": [
    "fresh isolated copy of the same neutral Next.js and TypeScript fixture",
    "exact public task prompt preserved before launch",
    "one bounded contestant invocation",
    "10-minute hard operator wall",
    "objective validation performed only after the contestant workspace was frozen"
  ],
  "treatment_variable": "None; this was a single-system capability test",
  "contestant_retries": 0,
  "contestant_repair_passes": 0,
  "post_submission_source_modifications": 0
}

Exact prompt

Replace the starter page with a responsive Issue Inbox interface. Show six seeded issues with realistic short titles. Add a search input that filters the visible issues by title as the user types. Show a live 'Showing X of 6 issues' count that updates with the filter. Include a Clear button that resets the search and restores all six issues. Make the interface usable on desktop and mobile. Use the existing Next.js and TypeScript setup. Do not add external services or unnecessary dependencies.

Supporting evidence

{
  "result": "PASS",
  "runtime": "4 minutes 46.944 seconds",
  "hermes_exit_code": 0,
  "production_build": "PASS",
  "interaction_checks": {
    "six_initial_issues": "PASS — 6 visible; Showing 6 of 6 issues",
    "search_filtering": "PASS — dark mode produced one matching issue",
    "live_count": "PASS — Showing 1 of 6 issues",
    "clear_restoration": "PASS — input cleared and all 6 issues returned"
  },
  "successful_model_calls": 36,
  "logged_cumulative_token_sums": {
    "input": 996393,
    "output": 9803,
    "total": 1006196,
    "label": "Sum of token values recorded on each successful model-call log line; not unique prompt size"
  },
  "http_429_responses": 0,
  "provider_backoff_seconds": 0,
  "monetary_cost": "Unavailable; no estimate made",
  "frozen_source_sha256": "1683e60039d32b055e6d982cdf0402d34e5a1926b8bed6ec522cbb5a67db2281",
  "contestant_source_repaired_after_run": false
}

Provenance

{
  "source_run": "runs/20260830T000508Z",
  "source_results": "runs/20260830T000508Z/results.json",
  "source_publication_analysis": "runs/20260830T000508Z/publication.json",
  "source_results_sha256": "438dd54d7c80d886d7b1804606eed2722184e2617401f2df28f618164d44217d"
}

Limitations

This is one small coding task and one contestant invocation, not a general benchmark or a broad comparison with other models.
The interaction checks establish the requested behavior for this preserved build; they do not measure long-term maintainability or performance across a larger application.
The token figures are cumulative values summed from the 36 logged model-call lines, not a unique prompt-size measurement or a provider-issued run-level usage object.
OpenRouter did not expose a run-level monetary cost in the preserved evidence, so no cost estimate is reported.