MiniMax Unmasks Space Bunny: M3.1 Flash Enters the Flash-Tier War With 200 Trillion Tokens of Stealth Data

MiniMax Unmasks Space Bunny: M3.1 Flash Enters the Flash-Tier War With 200 Trillion Tokens of Stealth Data

On September 23, 2026, a model with no name, no paper, and no press release appeared on OpenRouter. Developers called it Space Bunny. It was fast, cheap, and surprisingly capable — and for nearly four weeks, nobody outside MiniMax knew what it was. On October 19, the Shanghai-based lab pulled back the curtain: Space Bunny was MiniMax M3.1 Flash, a 428-billion-parameter multimodal Mixture-of-Experts model that had been quietly processing real developer traffic at scale before its official launch.

The reveal was more than a product announcement. It was a statement about how MiniMax thinks model releases should work — and a direct challenge to DeepSeek, which currently dominates the Flash-tier market with its V4.1-Flash-1017 checkpoint. Whether M3.1 Flash can hold that ground depends on how developers weigh its genuine strengths against a reliability problem that independent benchmarkers have been quick to flag.

The Space Bunny Strategy: 200 Trillion Tokens Before Launch

Most AI labs announce a model, then scramble to fix the bugs that real-world usage reveals. MiniMax inverted that sequence. According to Unite.AI's coverage of the launch↗, Space Bunny operated anonymously on OpenRouter for approximately 25.8 days, during which it processed over 200 trillion tokens — a volume that would take most labs months to accumulate through conventional beta programs.

The data MiniMax collected during that stealth run was not decorative. CellCog's analysis of the Space Bunny period↗ documents the concrete improvements that real-world feedback enabled before the official launch:

• Time to first token dropped from 1.95 seconds to 1.22 seconds — a 38% reduction in perceived latency
• Throughput increased from 70 to 100 tokens per second — a 45% gain in generation speed
• Error rate fell from 1.4% to 0.7%, halving the frequency of failed API calls

These are not marginal improvements. For developers building agentic pipelines where latency compounds across dozens of tool calls, a 38% reduction in time-to-first-token meaningfully changes the economics of a workflow. The stealth approach also gave MiniMax something more valuable than internal test data: it gave them a model that had already been stress-tested by the global developer community under production conditions.

The Space Bunny strategy represents a fundamental rethink of what a "launch" means. By the time MiniMax announced M3.1 Flash, the model had already survived the chaos of real-world deployment — something no internal benchmark can replicate.

Technical Architecture: What M3.1 Flash Actually Is

MiniMax M3.1 Flash is a multimodal Mixture-of-Experts model with 428 billion total parameters and approximately 23 billion active parameters per forward pass — the same parameter counts as its predecessor, the MiniMax M3. The architecture retains sparse attention and a native visual encoder, but the post-training work that produced M3.1 Flash from the M3 base is where the meaningful changes live.

Core Specifications

According to MarktechPost's release coverage↗ and the official MiniMax platform release notes↗, the model's key technical parameters are:

• Total parameters: ~428 billion (MoE architecture)
• Active parameters per token: ~23 billion
• Context window: 1,000,000 tokens
• Supported input modalities: Text, image, and video (up to 512MB video files)
• Output modality: Text only
• Reasoning: Always-on, with five selectable effort levels (Low, Medium, High, Xhigh, Max)

The always-on reasoning is the most operationally significant design choice. Unlike DeepSeek V4.1 Flash, which allows developers to disable thinking mode entirely for low-latency classification tasks, M3.1 Flash forces reasoning to remain active at all times. Attempts to disable it via the API return a 400 error. Developers can tune the cost-latency tradeoff by selecting effort level — "Low" can be substantially faster and cheaper than the default "Max" — but they cannot opt out of the reasoning process entirely.

Multimodal Differentiation

One area where M3.1 Flash clearly outpaces its primary competitor is native video support. Myclaw.ai's head-to-head comparison↗ confirms that DeepSeek V4.1 Flash handles text and images but does not natively ingest video files. M3.1 Flash accepts video inputs up to 512MB within its million-token context window — a capability that opens use cases in content moderation, video summarization, and multimodal agent workflows that DeepSeek's Flash tier simply cannot address.

Benchmark Performance: Strong on Agentic Tasks, Troubled on Factual Reliability

MiniMax's internal benchmark numbers are competitive. The Shattered.io launch analysis↗ and MarktechPost's coverage↗ report the following company-reported scores:

• DeepSWE v1.1 (software engineering): 73.7%
• OSWorld V2 (computer use/agentic): 61.7
• Automation-Bench v1.0.6: 62.2% — first among Chinese models
• BankerToolBench (finance/tool use): 76.2
• CyberGym (cybersecurity): 82%
• BabyVision (visual reasoning): 91%
• Agents' Last Exam (ALE): 29.6 — first among Chinese models

These numbers position M3.1 Flash as a serious contender for agentic and automation workloads. The Automation-Bench and ALE scores in particular suggest the model has been specifically optimized for the kind of multi-step, tool-using tasks that define modern AI agent deployments.

The Hallucination Problem

Independent evaluation tells a more complicated story. Artificial Analysis's assessment↗, reported by VentureBeat↗, places M3.1 Flash at an Intelligence Index score of 42–43 — a significant improvement over the M3's score of 29, but still behind DeepSeek V4.1 Flash 1017's score of approximately 45.

More concerning is the AA-Omniscience benchmark result. As Shattered.io's analysis↗ explains, the model achieved 56% raw factual accuracy — respectable for a Flash-tier model — but exhibited a 93% hallucination rate in failure states. This means that when M3.1 Flash does not know the correct answer, it provides a confidently wrong response 93% of the time rather than admitting uncertainty. The model's overall AA-Omniscience score was 16 as a result.

The 93% hallucination rate is not a death sentence for M3.1 Flash — it is a deployment constraint. For tasks with automated verification loops (code execution, structured data extraction, tool-call validation), the model's tendency to guess confidently is manageable. For open-ended knowledge retrieval without guardrails, it is a liability.

This distinction matters enormously for how developers should architect systems around M3.1 Flash. The model is well-suited to agentic pipelines where outputs are verified by execution — a code snippet either runs or it doesn't, a SQL query either returns results or it errors. It is poorly suited to tasks where factual accuracy must be taken on faith.

Pricing and the Flash-Tier Economics

MiniMax M3.1 Flash and DeepSeek V4.1 Flash share identical base API pricing: $0.15 per million input tokens and $0.60 per million output tokens. At this level, both models are dramatically cheaper than Western frontier alternatives. But the pricing structures diverge in ways that matter for production deployments.

Pricing Structure Comparison

According to Sandbase's detailed comparison↗ and kie.ai's analysis↗:

• MiniMax M3.1 Flash: Flat-rate pricing at $0.15/$0.60 regardless of time of day; cached input at $0.015/M tokens
• DeepSeek V4.1 Flash: Off-peak rates match at $0.15/$0.60, but peak weekday windows (01:00–04:00 and 06:00–10:00 UTC) double to $0.30/$1.20; more aggressive prompt-caching margins for stable system instructions

For teams running global, 24/7 agentic workloads, MiniMax's flat-rate structure eliminates the need to schedule around DeepSeek's peak pricing windows. For teams with predictable usage patterns and large recurring context blocks, DeepSeek's caching economics may still win on total cost.

MiniMax also claims the model delivers approximately 3x the inference speed of its M3 predecessor, with median output speeds exceeding 170 tokens per second and peak bursts over 200 tokens per second. Independent testing by Artificial Analysis measured 93–105 tokens per second — slower than MiniMax's claims, but still competitive for a model at this price point.

Open Weights: Scheduled but Not Yet Delivered

MiniMax announced that model weights would be released on October 23, 2026, allowing independent verification and community deployment. As of the time of writing, the specific license terms had not been disclosed — a notable gap given that license terms increasingly determine whether enterprise teams can legally deploy open-weight models in production.

The open-weight release, if it arrives on schedule, would give M3.1 Flash a significant advantage over its closed-weight competitors. The model's 428B parameter scale means local deployment requires substantial multi-GPU infrastructure — similar to the 8x B200 reference configuration used for M3 — making API consumption more practical for most teams. But for organizations with the hardware and the compliance requirements that make self-hosting necessary, the weights matter.

By contrast, DeepSeek V4.1 Flash 1017 already has its weights available on Hugging Face↗ under an MIT license — one of the most permissive licenses available. MiniMax's undisclosed license terms introduce uncertainty that some enterprise legal teams will find disqualifying until the terms are published.

MiniMax's Trajectory: From Video to Frontier Agentic Models

MiniMax is one of China's less-discussed frontier labs outside specialist circles, but its technical trajectory over 2026 has been consistently impressive. The company launched the MiniMax M3 flagship in June 2026, followed by the H3 multimodal video model in July — a model that drew significant attention for its video generation quality. M3.1 Flash represents the lab's push into the high-volume, cost-sensitive agentic tier that DeepSeek has dominated since the V4-Pro launch in August.

The Space Bunny stealth strategy also reveals something about MiniMax's competitive philosophy. Rather than racing to announce first and fix later, the lab chose to validate at scale before claiming the spotlight. That discipline — unusual in a sector where announcement velocity is often treated as a proxy for capability — suggests a team that is more interested in winning production deployments than winning press cycles.

Competitive Landscape

The Flash-tier market as of late October 2026 now has three serious Chinese contenders:

• DeepSeek V4.1 Flash 1017 — leads on coding benchmarks (DeepSWE 77.0 vs. 73.7), faster raw throughput (~209 tokens/sec), MIT-licensed weights already available, but peak-hour pricing doubles costs and no native video support
• MiniMax M3.1 Flash — leads on automation benchmarks (Automation-Bench 62.2 vs. 60.5), native video ingestion, flat-rate pricing, but weights pending and a 93% hallucination rate in failure states
• Kimi K3 (Moonshot AI) — the largest open-weight model at 2.8 trillion parameters, optimized for long-horizon agentic tasks, but at a different price and infrastructure tier than Flash models

For most developers choosing between DeepSeek and MiniMax at the Flash tier, the decision will come down to use case: coding-heavy pipelines favor DeepSeek's benchmark lead and MIT license; automation and multimodal workflows favor MiniMax's video support and flat-rate pricing.

Practical Takeaways for Developers

For teams evaluating M3.1 Flash for production use, the key operational considerations are:

• Agentic and automation workloads with verification loops are the model's sweet spot — the always-on reasoning and strong Automation-Bench scores make it well-suited to multi-step tool-using agents where outputs are validated by execution
• Fact-critical applications without guardrails should wait for the hallucination rate to be addressed or implement retrieval-augmented generation (RAG) and human-in-the-loop verification before deploying
• Video-processing pipelines have no comparable Flash-tier alternative from Chinese labs — M3.1 Flash's native 512MB video ingestion is a genuine differentiator
• Budget-sensitive global deployments benefit from flat-rate pricing that eliminates the need to schedule around DeepSeek's peak-hour windows
• Open-weight deployments should wait for the October 23 weight release and license disclosure before committing to infrastructure investment

The MiraFlow explainer↗ recommends that teams building on M3.1 Flash implement retrieval-augmented systems and structured output validation as standard practice — not as a workaround for the hallucination rate, but as good engineering hygiene for any model operating at this capability tier.

MiniMax M3.1 Flash is available now via the MiniMax API, the M Plan subscription, and the MiniMax Code desktop application. Open weights are scheduled for October 23, 2026, pending license disclosure. The model's stealth run as Space Bunny may be over, but the real test — whether it can displace DeepSeek in production agentic pipelines — is just beginning.