Chinese AI startup MiniMax has officially launched M3.1 Flash, an agentic model the company says is roughly three times faster and 70 percent cheaper to run than its predecessor. The announcement, published on MiniMax's website on October 19 and picked up the same day by Reuters, also ends one of the more unusual mysteries in the AI community: the anonymous model known as "Space Bunny," which quietly processed more than 200 trillion tokens while operating under a codename, was an early version of MiniMax M3.1 Flash.

The Space Bunny Reveal For more context on this story, see our ongoing latest AI developments.

For weeks, developers had been testing a capable model listed under the Space Bunny name without any clear indication of who built it. MiniMax's launch post confirmed that Space Bunny was, in fact, an early version of M3.1 Flash that had been serving traffic anonymously. The company says the model accumulated more than 200 trillion processed tokens during that period, giving it an unusually large real-world evaluation base before its official debut.

The strategy mirrors a broader pattern among Chinese AI labs, which have increasingly used anonymous or preview releases to gather feedback before putting a model's real name on it. MiniMax itself previewed the model in late September as M3.1-Flash-Preview, with no benchmarks and no pricing, as AI Buzz Wire reported at the time.

Benchmarks Aimed at Agents, Not Just Chat

MiniMax positions M3.1 Flash as a next-generation "Flash" model built for agentic work — computer use, coding, cybersecurity and long enterprise workflows — rather than conversational chat alone. According to the company's published results:

  • BankerToolBench: 76.2 — the company says the model can complete real enterprise workflows across SaaS systems end to end, spanning sales, marketing, operations, customer service, finance and HR tasks, with strengths in multi-step planning, document consistency, financial compliance and chained tool calls.
  • DeepSWE v1.1: 73.7% — placing it, by MiniMax's account, alongside top models on complex software engineering tasks that require independent code comprehension, feature development, bug fixing and test validation.
  • OSWorld V2: 61.7 and Agents' Last Exam (ALE): 29.6 — benchmarks that measure computer operation and performance on real professional tasks. MiniMax says M3.1 Flash ranks first among Chinese models on both.
  • CyberGym: 82% — which the company cites as evidence of leading capability in vulnerability discovery and validation, alongside internal evaluations in code auditing, penetration testing, malware analysis and reverse engineering.

As always with vendor-reported benchmarks, independent verification remains limited, and leaderboard comparisons depend heavily on configuration. MiniMax says results have been optimized through "more aggressive sparsity" in the inference stack.

Speed and Pricing: The Aggressive Part

The economic claims are arguably the bigger story. MiniMax says everyday output speeds reach 200+ tokens per second, with median (P50) throughput consistently above 170 tokens per second — and that overall usage cost for agentic tasks is 70 percent lower than the model's predecessor.

Pricing is flat per million tokens:

| Tier | Price (USD per 1M tokens) |
|---|---|
| Input | $0.15 |
| Output | $0.60 |
| Cached input | $0.015 |

That undercuts most frontier-class API pricing by a wide margin and continues the price war that has defined the agentic model market through 2026. It arrives days after DeepSeek released its V4.1-Flash-1017 update under an MIT license with open weights, keeping pressure on every pricing tier of the market. For high-volume agent workloads — where a single enterprise workflow can involve hundreds of model calls — the spread between $0.60 and multi-dollar output pricing compounds into the difference between a viable product and an unaffordable one.

The speed claims matter for a different reason. Computer-use agents stall when each click-and-observe step waits seconds for a response; throughput above 170 tokens per second at the median makes multi-step GUI automation feel responsive enough for production use. MiniMax is betting that agent economics, not raw benchmark leadership, is what converts enterprise pilots into deployments.

MarkTechPost, which covered the launch, described M3.1 Flash as a 428-billion-parameter mixture-of-experts model, though MiniMax's own announcement does not state a parameter count.

Availability: API Now, Weights Later

M3.1 Flash is available immediately through the MiniMax API, the company's M Plan subscription, and MiniMax Code, its desktop coding agent. Model weights will be released later, the company said, and a more capable "Pro" variant is also on the roadmap. No dates were given for either.

The staged approach — API first, open weights later — has become MiniMax's standard playbook. The company released its H3 video model with open weights in July, and its enterprise revenue nearly quadrupled year over year to $116.6 million in the first half of 2026.

What It Means for Developers

For teams building agent pipelines, M3.1 Flash's pitch is straightforward: near-frontier agentic performance at commodity prices, with throughput numbers that make multi-step computer-use workflows economically viable. The open questions are how the vendor-reported benchmarks hold up under independent testing, and how quickly the promised weight release follows.

With DeepSeek, Z.ai and now MiniMax all shipping agentic models at aggressive price points within the same month, the competitive gap between Chinese open-weight ecosystems and Western closed APIs continues to narrow — and buyers are the immediate beneficiaries.

_This article is based on MiniMax's official launch announcement (October 19, 2026), reporting by Reuters, and coverage by MarkTechPost._

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →