MiniMax M3.1 Flash in practice: tests and comparisons with DeepSeek V4.1 Flash

Hands-on test: a single direct API call at max thinking with 50K tokens and no harness or tool calls produced the most visually impressive pagoda so far, better than DeepSeek V4.1 Flash.

MiniMax confirmed that OpenCode's free Space Bunny model is M3.1 Flash, offering $0 input/output, zero data retention, 1M context and image/video input, plus how to enable it.

The author recalls testing Space Bunny free in OpenCode Zen on Oct 2 — fast, finding 12/12 keys in roughly 1M characters of context — and notes MiniMax has now confirmed it as M3.1 Flash.

The stealth model Space Bunny in OpenCode turned out to be MiniMax-M3.1-Flash, which is still free to test there.

The author tries M3.1 Flash through the API and finds it notably fast, with a video of the output.

A user recalls that free Space Bunny was the most-used model on OpenCode and OpenRouter, with speed near DeepSeek, and hopes the official release keeps the price and ships open weights.

MiniMax's official account posts a teaser video for the upcoming M3.1 Flash, calling back to Space Bunny by name.

A user describes their OpenCode setup after replacing Muse Spark 1.3: GLM 5.3 Flash for planning, DeepSeek V4.1 Flash for building (~8 hours of free quota), Space Bunny's 1M window when context runs out, and MiMo 2.6 Flash as backup.

MiniMax is retiring M2.5 and M2.7 on November 18, 2026; from October 19, M Plan and Token Plan requests auto-route to M3.1 Flash Preview with model names and allowances unchanged.

After Muse Spark 1.3 left OpenCode's free lineup, the author suggests GLM 5.3 Flash for planning, DeepSeek V4.1 Flash for coding, and Space Bunny (the anonymous M3.1 Flash) for 1M-context work where DeepSeek's 200K compacts too early.

MiniMax's official API model list now includes minimax-m3.1-flash, with a screenshot showing the new entry.

On OpenCode Go, space-bunny, gpt-6.1-sol and claude-sonnet-5-5 all passed 15/15 runs of 5 small hidden-test coding tasks, at $0.022, $0.223 and $2.435 API-equivalent cost; the tasks were too easy to separate them.