{"slug": "minimax-quietly-launches-m3-1-flash-after-weeks-as-a-mystery-space-bunny-model", "title": "MiniMax quietly launches M3.1-Flash after weeks as a mystery Space Bunny model", "summary": "MiniMax launched M3.1-Flash-Preview inside its MiniMax Code agent tool on September 27, 2026, with no model card, benchmark report or public pricing, after developers spent two weeks testing the same system as a free OpenRouter model called Space Bunny Alpha under the provider tag \"Stealth,\" according to CodersEra reporting that matched parameter fingerprints and behavior to MiniMax's model family. The stealth listing, which offered a 1 million token context window and video input at no cost, expired on October 5. MiniMax's full M3 model is priced at $0.30 per million input tokens and $1.20 per million output tokens with an 80.5% score on SWE-bench Verified, and the quiet drop follows similar no-fanfare releases from DeepSeek (V4.1-Flash-1017) and Nvidia (Nemotron 3.5 Bolt) as Chinese labs compete on price and self-reported benchmarks for agentic coding work.", "body_md": "*MiniMax spent weeks letting developers test its newest model under a fake name before admitting it built it. That trick is becoming the normal way Chinese AI labs launch now.*\n\nOn September 23, a free model called Space Bunny Alpha showed up on OpenRouter under the mysterious provider tag \"Stealth.\" It had a 1 million token context window, could read video, and cost nothing to use. Developers on OpenRouter and in coding tool communities spent the next two weeks poking at it, comparing outputs, and guessing who actually built it. The listing quietly expired on October 5. By then the consensus favorite was MiniMax, and the guess turned out right: Space Bunny was MiniMax's M3.1-Flash, according to reporting from CodersEra that tracked the model's parameter fingerprints and behavior against MiniMax's known model family.\n\nMiniMax never put out a press release. Instead, on September 27 the company slipped \"M3.1-Flash-Preview\" directly into MiniMax Code, its own coding tool, available to anyone already on a Token Plan subscription. No launch event, no blog post explaining the architecture, just a model that appeared where paying users would find it. That's the same quiet-drop playbook DeepSeek used for V4.1-Flash-1017 two days ago and Nvidia used rolling out Nemotron 3.5 Bolt. Ship first, let the benchmarks and the Reddit threads do the talking later.\n\nThe bet is the same across all three labs: agentic coding work, the kind where a model calls tools, edits files, and runs tests in a loop, burns tokens fast. So the model that wins isn't the smartest one. It's the cheapest one that's still good enough. MiniMax's full M3 already undercuts that math aggressively: $0.30 per million input tokens, $1.20 per million output tokens, and an 80.5% score on SWE-bench Verified that MiniMax has used to argue M3 beats GPT-5.5 and Gemini 3.1 Pro at a fraction of their cost. M3.1-Flash is meant to go further down that same curve. Faster responses, a smaller footprint, tunable reasoning effort so a developer can dial compute up or down depending on whether the task is a quick lint fix or a multi-file refactor.\n\nThat's a direct shot at DeepSeek's turf. DeepSeek has spent the better part of 2026 racing out flash variants and cutting prices to hold onto developer mindshare, and it dropped V4.1-Flash-1017 on Hugging Face just five weeks after the base V4.1-Flash model, with no fanfare either. Nvidia's Nemotron 3.5 Bolt is chasing the same agentic coding niche from the hardware-vendor side. Between them, MiniMax, Alibaba's Qwen, DeepSeek, Zhipu's GLM, and ByteDance's Doubao now ship dozens of models that compete head to head with Western frontier labs. When nearly everyone can serve a frontier-class model cheaply, price stops being the thing that sets a lab apart. Benchmark leadership becomes the battleground instead.\n\n[MiniMax slips a new coding model into its agent tool without a price tag](https://startupfortune.com/minimax-slips-a-new-coding-model-into-its-agent-tool-without-a-price-tag/)\n\nMiniMax launched M3.1-Flash-Preview inside its MiniMax Code agent tool on September 27, 2026, with no model card, no benchmark report and no public pricing. The release follows community fingerprinting that linked a free OpenRouter stealth model, Space Bunny Alpha, to the same system days earlier. - [MiniMax coding model M3.1 Flash Preview release](https://startupfortune.com/minimax-slips-a-new-coding-model-into-its-agent-tool-without-a-price-tag/) - [new AI coding agent tool MiniMax Code launch](https://startupfortune.com/minimax-slips-a-new-coding-model-into-its-agent-tool-without-a-price-tag/)\n\nWhich is exactly where the trouble starts. Benchmark numbers in this race are almost entirely self-reported. And 2026 has already produced a cautionary case: when Moonshot AI published Kimi K2's score on the Humanity's Last Exam benchmark at roughly 50%, independent testers who re-ran the same evaluation got 29.4%, a gap north of 20 points. Nobody has caught MiniMax doing anything like that with M3.1-Flash specifically, and to be fair to the company, it hasn't actually published a flashy benchmark claim for this release yet. According to tracking from BenchLM, there's no widely published, weighted benchmark score for M3.1-Flash-Preview as of this week. One early private evaluation from ModelDial put the model at just 30 out of 100 on a backend-and-testing coding task at low reasoning effort, a result nowhere near what MiniMax touts for full M3.\n\nThat gap matters because it's the whole point of releasing a stealth model before a named one. Space Bunny let real developers stress-test M3.1-Flash on real tasks for two straight weeks, with nobody's reputation or marketing budget riding on the result. By the time MiniMax folded it into MiniMax Code under its real name, the model had already earned whatever reputation it was going to earn organically: through actual usage, not a benchmark chart timed to a launch.\n\nFrankly, that's smarter than what most labs do. A model that has to survive two weeks of anonymous scrutiny before its maker claims it says more than a cherry-picked SWE-bench number released the same day as the announcement. Whether M3.1-Flash actually holds up against DeepSeek's and Nvidia's flash models once independent benchmarks catch up is still an open question. MiniMax clearly didn't want to answer it with a press release. It wanted the market to answer it first.\n\n**Also read:** [Elon Musk Pledges a SpaceX Share to Every American Kid Next Year](https://startupfortune.com/elon-musk-pledges-a-spacex-share-to-every-american-kid-next-year/) • [GPT-6 Astra Beats Claude Sonnet 5.5 on Science Tasks Using Far Fewer Tokens](https://startupfortune.com/gpt-6-astra-beats-claude-sonnet-55-on-science-tasks-using-far-fewer-tokens/) • [Cerebras stock sank below its IPO price as the AI selloff hit chips](https://startupfortune.com/cerebras-stock-sank-below-its-ipo-price-as-the-ai-selloff-hit-chips/)\n\n*This article is posted in [Technology News](https://startupfortune.com/category/technology/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/minimax-quietly-launches-m3-1-flash-after-weeks-as-a-mystery-space-bunny-model", "canonical_source": "https://startupfortune.com/minimax-quietly-launches-m31-flash-after-weeks-as-a-mystery-space-bunny-model/", "published_at": "2026-10-19 10:50:13+00:00", "updated_at": "2026-10-19 11:47:39.101684+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-products", "ai-tools", "generative-ai"], "entities": ["MiniMax", "M3.1-Flash-Preview", "MiniMax Code", "Space Bunny Alpha", "OpenRouter", "DeepSeek", "Nvidia", "Moonshot AI"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/minimax-quietly-launches-m3-1-flash-after-weeks-as-a-mystery-space-bunny-model", "markdown": "https://wpnews.pro/news/minimax-quietly-launches-m3-1-flash-after-weeks-as-a-mystery-space-bunny-model.md", "text": "https://wpnews.pro/news/minimax-quietly-launches-m3-1-flash-after-weeks-as-a-mystery-space-bunny-model.txt", "jsonld": "https://wpnews.pro/news/minimax-quietly-launches-m3-1-flash-after-weeks-as-a-mystery-space-bunny-model.jsonld"}}