MiniMax Quietly Launched a New Coding Model — M3.1-Flash-Preview

 

 

MiniMax has a habit of shipping first and explaining later - and on September 27, 2026, the Chinese AI lab did it again. A new coding model, M3.1-Flash-Preview, quietly went live inside MiniMax Code, the company's own coding agent product. There was no model card, no benchmark table, no pricing page, and no public API. Just a new model, available to developers, announced in a single low-key post.

According to MiniMax's own announcement on X, the model "debuts today on MiniMax Code" and is "built for everyday development, fast, reliable, and ready for real work, from quick bug fixes to full features." That is effectively the entire launch. For a frontier-class coding model, the restraint is remarkable - and it appears to be deliberate.

What is public so far: the model supports up to a 1M token context window, and MiniMax Code now exposes five reasoning-effort tiers - low, medium, high, xhigh, and max - giving developers fine-grained control over the tradeoff between speed and depth of reasoning. Beyond that, details are thin: no published eval scores, no architecture notes, and no word on when (or whether) it will land on a public API.

The quiet launch follows a curious prelude. Four days earlier, an anonymous free model appeared on OpenRouter under the name "Space Bunny Alpha." Developers quickly got to work fingerprinting it: tokenizer and error-behavior tests returned striking matches - 24/24 and 50/50 - leading the community to conclude it was an early build of the same MiniMax system. In other words, the model appears to have been benchmark-tested in the wild under a pseudonym before its official debut. MiniMax has not confirmed the connection, but the fingerprints are hard to ignore.

This is becoming a pattern. MiniMax's previous coding model, M3, scored 80.5% on SWE-bench Verified and outperformed GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro - at a fraction of the cost, with M3 API pricing at just $0.30 per million input tokens and $1.20 per million output tokens. The lab has consistently let results speak louder than marketing.

Why it matters: the coding-model race is shifting from benchmark one-upmanship to iteration velocity. Shipping a model directly into a first-party coding agent - gathering real developer feedback before publishing a model card - compresses the loop between release and improvement. If M3.1-Flash-Preview follows the M3 playbook, expect the paperwork (benchmarks, pricing, API access) to arrive once the model has already proven itself in production.

Our take: iteration speed is the product. In a market where every lab claims frontier performance, MiniMax is betting that shipping first and documenting later is the faster route to developer trust. Stealth launches aren't accidents anymore - they're strategy.