MiniMax M3.1 Flash Coding Tests: Vendor Scores vs What We Measured

Vendor-reported scores, not Spacebunny scores

MiniMax’s research post dated 19 October 2026 lists the figures below. Spacebunny did not run DeepSWE, OSWorld, BankerToolBench, Agents’ Last Exam, or CyberGym. Do not treat this table as an independent leaderboard.

• BankerToolBench: 76.2. Vendor-reported.
• DeepSWE v1.1: 73.7%. Vendor-reported. The same day’s official X post rounded coding to 74%.
• OSWorld V2: 61.7. Vendor-reported.
• Agents’ Last Exam: 29.6. Vendor-reported. The X post rounded this to 30%.
• CyberGym: 82%. Vendor-reported.
• Speed: MiniMax says everyday output can pass 200 tokens per second, with a median around 170. That is vendor-reported, not a Spacebunny timing.

What OpenRouter published separately

On 21 October 2026 the OpenRouter page for minimax/minimax-m3.1-flash:free showed a provider throughput around 84 tokens per second and latency around 1.51 seconds (their P50). That is OpenRouter’s figure for the free route. It is not the 200 tokens per second MiniMax quotes, and it is not a measurement from this site.

Independent test status

This revision does not publish an independent score or a model transcript. The OpenRouter key lives only as a Worker secret, so this page was not written from a fresh completion. Inventing an output would be worse than leaving the cell blank.

If you want a reproducible check inside our limits, use the playground on or after 21 October 2026 with model id minimax/minimax-m3.1-flash:free, temperature 0.2, and this prompt. The playground still caps a reply at 2048 tokens and a request near 100k tokens, so the result is not comparable to DeepSWE.

Write a TypeScript function sumPositive(nums: Array<number | null>): number that adds the numbers and treats null as 0. Return only the function. Do not use any or as.

How to read a result if you run it

• Record the date, the model id shown in the playground, and the full reply.
• A correct function returns 0 for [], 0 for [null], and 3 for [1, null, 2].
• One sample is not a benchmark. It does not support a percentage score.
• Vendor exams stay in the first section until someone publishes a rerun with the prompt set, harness, and date.