MiniMax M3.1-Flash-Preview: The Quiet Launch, Stealth Benchmarking, and the New Speed Tier for Everyday Coding

Executive Summary

On September 27, 2026, Shanghai-based MiniMax quietly launched M3.1-Flash-Preview — a new frontier coding model designed for "everyday development, from quick bug fixes to full features." Unlike the company's previous flagship launches, this release came with no model card, no published benchmarks, no public pricing, and no direct API endpoint. The model is available exclusively through MiniMax Code, the company's open-source coding agent, and through the MiniMax Token Plan.

The launch follows a now-familiar pattern in the Chinese AI ecosystem: a stealth benchmark run under the anonymous name "Space Bunny Alpha" on OpenRouter four days earlier, followed by a quiet in-product launch. Developers fingerprinted the anonymous model through tokenizer comparisons and error-behavior tests, confirming it matched MiniMax's model family with 100% accuracy across 50 probes.

M3.1-Flash-Preview builds on the M3 architecture — a 428-billion-parameter MoE model with 23 billion active parameters per token, MiniMax Sparse Attention (MSA), and a 1 million-token context window. Key differentiators include forced reasoning (thinking cannot be disabled), a six-tier effort ladder (default, low, medium, high, xhigh, max), and dramatically faster inference than M3 on routine tasks — completing simple questions in 1.02 seconds versus M3's 1.92 seconds, and coding tasks 2× faster while using far fewer reasoning tokens.

The release reflects a strategic shift: rather than publishing benchmark charts and pricing sheets, MiniMax is letting the model speak for itself through developer experience, iterating at a pace that makes switching costs the primary competitive barrier.

---

1. The Stealth Launch Pattern

1.1 What Happened

The M3.1-Flash-Preview launch unfolded in three phases:

1. September 23: An anonymous model called "Space Bunny Alpha" appeared on OpenRouter under the identifier stealth/space-bunny-alpha, priced at $0 for both input and output during its preview window.
2. September 23–26: Developers independently fingerprinted the model through tokenizer comparisons and error-behavior tests. An independent test sent 24 token-count probes through OpenCode's Space Bunny route and found all 24 matched MiniMax models; a separate measurement set confirmed 50 out of 50 matches.
3. September 27: MiniMax's official @MiniMax_Agent account posted on X: "M3.1-Flash-Preview debuts today on MiniMax Code. Built for everyday development — fast, reliable, and ready for real work, from quick bug fixes to full features." That was the entire announcement.

1.2 Why This Matters

This pattern — free anonymous benchmark run followed by quiet in-product launch — is becoming a standard practice in the Chinese AI ecosystem. It serves multiple purposes:

• Real-world data collection: Gather performance data from actual developers without the pressure of published benchmarks.
• Risk mitigation: If the model underperforms, there's no published claim to retract.
• Developer adoption: Let the model's quality drive adoption organically rather than through marketing.
• Iteration speed: Ship fast enough that competitors can't catch up before the next release.

This contrasts sharply with the Western approach, where labs typically publish model cards, benchmark tables, and blog posts before letting anyone touch a new release.

1.3 The MiniMax Code Vehicle

M3.1-Flash-Preview launched inside MiniMax Code, the company's open-source coding agent, which was itself open-sourced under MIT on September 18, 2026 — just nine days before the model drop. The tool provides three entry points:

Entry Point	Purpose
mcode	Interactive terminal UI
mcode exec	Script-oriented execution
mcode acp	Editor integration (ACP protocol)

All three share one in-process runtime rather than shelling out to a server, enabling tight integration with the model's reasoning capabilities.

---

2. Model Specifications

2.1 Technical Details

Property	Value
Model name	MiniMax-M3.1-Flash-Preview
Architecture	MoE with MiniMax Sparse Attention (MSA)
Total parameters	~428B (inherited from M3 family)
Active parameters	~23B per token
Context window	1,000,000 tokens (default)
Maximum output	128,000 tokens
Input modalities	Text, Image, Video
Output modalities	Text
Reasoning	Forced on (cannot be disabled)
Effort levels	Default, Low, Medium, High, XHigh, Max (6 tiers)
API endpoints	chatcompletion_v2, Anthropic-compatible (/anthropic/v1)
Availability	MiniMax Code, Token Plan only

2.2 The Six-Tier Effort Ladder

The effort ladder is a significant differentiator. Most reasoning models offer three tiers (low, medium, high). M3.1-Flash-Preview offers six, suggesting MiniMax is tuning this specifically for developers who want to dial cost and speed against task difficulty on the fly:

Effort Level	Use Case	Typical Thinking Tokens
Default	General purpose	~50-100
Low	Quick bug fixes, trivial questions	~15
Medium	Standard coding tasks	~100-200
High	Complex features, refactoring	~300-600
XHigh	Multi-file changes, architecture work	~600-1200
Max	Full feature implementation, research	~1200+

Independent testing confirmed the effort ladder works through the Anthropic-compatible endpoint. On a small probe prompt, the thinking block grew from 129 characters at "low" to 365 characters at "max".

2.3 Forced Reasoning

Unlike M3, which allows users to disable thinking, M3.1-Flash-Preview forces reasoning on. This is a deliberate design choice — the model always thinks before answering, even for trivial questions. The difference is in how much it thinks, controlled by the effort level.

This aligns with the "Flash" positioning: the model thinks efficiently rather than optionally. On a one-line question, it spent ~15 tokens thinking and answered in 1.02 seconds — far less deliberation than M3 (26 tokens, 1.92 seconds) or M2.7-highspeed (103 tokens, 2.88 seconds).

---

3. Performance: Speed Where It Counts

3.1 Throughput Benchmarks

Independent testing across three task shapes reveals the Flash tier's speed advantage:

Task	M3.1-Flash-Preview	M3	M2.7-highspeed
Trivial question	1.02s	1.92s	2.88s
FizzBuzz	94.6 tok/s	71.0 tok/s	~50 tok/s
Dedupe script	4.26s	8.11s	32.97s

Key observations:

• Time to first token is close across all models — the wall-clock wins come from shorter deliberation, not faster first bytes.
• On the dedupe task, M3.1-Flash-Preview used 384 output tokens (213 reasoning) versus M3's 898 output tokens (622 reasoning) — the Flash model produces the same result with 57% fewer tokens.
• Against M2.7-highspeed (the previous "fast" tier), the new model is nearly 8× faster on the dedupe task.

3.2 Quality Assessment

Speed means nothing if the code is wrong. Independent testing across six tasks (CSV parser, pagination bug fix, async cache, rate limiter, typed result wrapper, SQL report) with hidden test suites found:

Task	M3.1-Flash-Preview	M3	M2.7-highspeed
CSV parser	Pass	Pass	Pass
Pagination fix	Pass	Pass	Fail
Async cache	Pass	Pass	Fail
Rate limiter	Pass	Pass	Pass
Typed wrapper	Pass	Pass	Fail
SQL report	Pass	Pass	Fail

The Flash model matched M3 on all six tasks while being significantly faster, and outperformed M2.7-highspeed on 4 of 6 tasks.

3.3 Real-World Behavior

Beyond benchmarks, two behaviors stand out from independent testing:

1. Self-correction: When its initial test run failed, the model traced the failure to itself — it had hand-miscounted its own fixture strings, verified the real values against coreutils wc and awk, and corrected the expectations. The implementation had been right all along.

2. Test discipline: Asked for three tests, it wrote four — keeping the error-path test separate — and offered to collapse to exactly three if insisted. This kind of judgment is what benchmark suites don't capture.

---

4. The M3 Foundation

4.1 Architecture

M3.1-Flash-Preview inherits its architecture from M3, launched in June 2026:

• 428B total parameters in a Mixture-of-Experts design
• 23B active parameters per token during inference
• MiniMax Sparse Attention (MSA) — a proprietary attention mechanism that reduces compute while maintaining performance
• 1M-token context window with guaranteed minimum of 512K tokens
• Native multimodality — text, image, and video input from step zero of pretraining

4.2 M3's Published Benchmarks

The M3 flagship has published benchmarks that establish the family's capability ceiling:

Benchmark	M3 Score	Context
SWE-bench Verified	80.5%	MiniMax-reported
SWE-Bench Pro	59.0%	Ahead of GPT-5.5 and Gemini 3.1 Pro per MiniMax
Terminal-Bench 2.1	66%	Agentic coding
BrowseComp	83.5	Autonomous browsing (surpasses Opus 4.7 at 79.3)
PostTrainBench	37.1	#3 overall (behind Opus 4.7 at 42.4 and GPT-5.5 at 39.3)

4.3 The M3 Family Lineup

Model	Positioning	Context	Reasoning	Effort Levels
M3.1-Flash-Preview	Fast everyday coding	1M default	Forced on	6 tiers
M3	Frontier coding & agentic	512K default	Optional	3 tiers
M2.7	Self-evolving agent	200K	Optional	Standard
M2.7-highspeed	Fast M2.7	200K	Optional	Standard

---

5. Pricing & Access

5.1 The Pricing Gap

M3.1-Flash-Preview has no published pricing. However, the M3 family's pricing establishes the context:

Model	Input Price	Output Price	Cache Read
M3	$0.30/M	$1.20/M	$0.06/M
M2.7	$0.30/M	$1.20/M	$0.06/M
Claude Sonnet 5	$2.00/M	$10.00/M	$0.20/M
Claude Opus 5	$5.00/M	$25.00/M	$0.50/M
GPT-6 Sol	$2.00/M	$10.00/M	—
Grok 4.7	$2.00/M	$6.00/M	$0.50/M

At M3's pricing, MiniMax undercuts Claude Sonnet 5 by 85% on input and 88% on output. The Flash tier, positioned as a faster variant, is expected to maintain similar or lower pricing.

5.2 Access Restrictions

Access Method	M3.1-Flash-Preview	M3
MiniMax Code	✅	✅
Token Plan	✅	✅
Public API (pay-as-you-go)	❌	✅
OpenRouter	❌	✅
Hugging Face weights	❌	Coming soon

The restriction to Token Plan and MiniMax Code is a strategic choice — it creates a walled garden where MiniMax controls the user experience and can gather usage data before a broader release.

5.3 The Token Plan Incentive

To drive adoption, MiniMax announced:

• Doubled daily check-in credits from September 28 to October 7 (UTC+8)
• Token Plan quotas reset at launch
• Credits work across the whole model catalogue, including H3 video generation

---

6. The API Reality: It Works Outside the Walled Garden

6.1 The Lock Isn't Actually a Lock

Despite the official announcement limiting access to MiniMax Code and Token Plan, independent testing found that M3.1-Flash-Preview resolves happily on public endpoints:

• The model ID MiniMax-M3.1-Flash-Preview works on the chatcompletion_v2 endpoint
• It also works on MiniMax's Anthropic-compatible endpoint (/anthropic/v1)
• The effort ladder works through the Anthropic-compatible endpoint using output_config: { "effort": "low" }

6.2 Integration Example

{
  "model": "MiniMax-M3.1-Flash-Preview",
  "max_tokens": 1024,
  "messages": [
    {"role": "user", "content": "Explain single-flight caching in two sentences."}
  ],
  "output_config": {"effort": "low"}
}

This works against https://api.minimax.io/anthropic/v1 with a valid MiniMax API key. The trailing /v1 is load-bearing — without it, every request 404s.

6.3 OpenCode Integration

For OpenCode users, a custom provider configuration enables the model:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "minimax": {
      "npm": "@ai-sdk/anthropic",
      "name": "MiniMax",
      "options": {
        "baseURL": "https://api.minimax.io/anthropic/v1",
        "apiKey": "{env:MINIMAX_API_KEY}"
      },
      "models": {
        "MiniMax-M3.1-Flash-Preview": {}
      }
    }
  }
}

---

7. The Bigger Picture: China's AI Model Flood

7.1 The September 2026 Landscape

M3.1-Flash-Preview is one release in an unprecedented wave of Chinese model launches:

Model	Release Date	Key Feature
DeepSeek V4.1-Flash	Sep 10	Causal encoder-decoder, 552B MoE, $0.30/M
Qwen-Image-2.1	Sep 20	7B open image model, native transparency
Step 5 Preview	Sep 21	600B sparse MoE, open-weight Pareto frontier
Grok 4.7	Sep 21	Long-running agents, $2/M
GPT-6 Sol/Luna	Sep 22	Three-tier strategy, 50% price cut
Claude Opus 5.5	Sep 22	Fable 5.1 performance at 40% less cost
MiniMax M3.1-Flash-Preview	Sep 27	Stealth launch, 6-tier effort, forced reasoning

7.2 Market Share Shift

According to CNBC's reporting on OpenRouter data, Chinese models accounted for 57% to 67% of total token usage on the platform for the week including September 14, up from just 6% to 13% in February. Vercel saw a similar jump, with Chinese models' share rising to 55% in August from 11% in January.

The driver is not mysterious: for agentic coding and customer-service workloads, Chinese open models run 60% to 90% cheaper than leading U.S. alternatives.

7.3 The Geopolitical Dimension

Washington is watching this shift with concern. Daniel Remler, a senior fellow at the Center for a New American Security, told CNBC that Chinese AI represents "real economic and security risks for the United States" and warned that integrating Chinese models could pull countries into China's technology sphere of influence.

Meanwhile, Huawei used its annual conference in Shanghai to unveil the Atlas 960 SuperPoD computing cluster, with Ascend chips increasingly part of enterprise procurement as Chinese AI companies push toward domestic silicon rather than Nvidia hardware.

---

8. Connection to Prior Research

8.1 The MiniMax Evolution

M3.1-Flash-Preview continues MiniMax's rapid iteration cycle:

Model	Release	Key Innovation
M2.7	July 2026	Self-evolving agent harnesses (Minimax M27 Self Evolving Agent Harness Open Weight Frontier 2026 07 16)
M3	June 2026	428B MoE, MSA architecture, 1M context, 80.5% SWE-bench Verified
M3.1-Flash-Preview	Sep 2026	Forced reasoning, 6-tier effort, stealth launch, 2× speed

8.2 Positioning in the Frontier Landscape

Model	Input Price	Output Price	Key Differentiator
M3.1-Flash-Preview	~$0.30/M	~$1.20/M	Speed + forced reasoning + 6-tier effort
Grok 4.7	$2/M	$6/M	Domain expertise, long-running agents
GPT-6 Sol	$2/M	$10/M	Mid-tier agentic coding
Claude Opus 5.5	$4/M	$20/M	Safety + efficiency
DeepSeek V4.1-Flash	$0.30/M	$1.20/M	Causal encoder-decoder architecture

See Grok 4 7 Long Running Agents Coding Knowledge Work New Safeguard Stack 2026 09 25 for Grok 4.7, Gpt 6 Sol Luna 50 Percent Price Cut Agentic Coding Cost Efficiency Frontier 2026 09 24 for GPT-6, Claude Opus 5 5 Fable 5 1 Performance 40 Percent Cost Reduction Pacing Frontier Safeguards 2026 09 23 for Claude Opus 5.5, and Deepseek V4 1 Flash Causal Encoder Decoder Kv Cache Compression Agentic Coding 2026 09 14 for DeepSeek V4.1-Flash.

8.3 The Stealth Benchmarking Pattern

The Space Bunny Alpha → M3.1-Flash-Preview pattern mirrors previous stealth launches in the Chinese AI ecosystem. This approach — collect real-world data anonymously before official launch — is becoming a standard practice that Western labs have not yet adopted.

---

9. Key Takeaways

1. Stealth launch as strategy. MiniMax skipped the model card, benchmarks, and pricing sheet — letting the model speak for itself through developer experience. This is becoming a standard pattern in the Chinese AI ecosystem.

2. Six-tier effort ladder. The six reasoning effort levels (default through max) are unique among frontier models, suggesting MiniMax is optimizing for developers who want fine-grained control over the cost-speed-quality tradeoff.

3. Forced reasoning. Unlike M3, thinking cannot be disabled. The model always thinks before answering, but the amount of thinking is controlled by the effort level.

4. 2× speed on routine tasks. On simple questions, M3.1-Flash-Preview answers in 1.02 seconds versus M3's 1.92 seconds, using 57% fewer tokens for the same result.

5. The lock isn't a lock. Despite official restrictions to MiniMax Code and Token Plan, the model works on public API endpoints and can be integrated into third-party tools like OpenCode.

6. Price remains the killer feature. At ~$0.30/M input and ~$1.20/M output, MiniMax undercuts Claude Sonnet 5 by 85% and GPT-6 Sol by 85% on input tokens.

7. Chinese model dominance is accelerating. Chinese models now account for 57-67% of OpenRouter usage, up from 6-13% in February, driven by 60-90% cost advantages.

8. Iteration speed as the product. MiniMax is shipping so fast that switching costs never build up — a strategy that Western labs' more deliberate release cycles can't match.

---

10. References & Resources

Official Sources

• MiniMax M3 Model Page — Official M3 specifications and capabilities
• MiniMax Platform Documentation — Model catalog and API documentation
• MiniMax Code Releases — Open-source coding agent releases
• MiniMax Agent X Account — Official announcement channel

Related Research in This Journal

• Minimax M27 Self Evolving Agent Harness Open Weight Frontier 2026 07 16 — MiniMax M2.7 self-evolution analysis
• Grok 4 7 Long Running Agents Coding Knowledge Work New Safeguard Stack 2026 09 25 — Grok 4.7 comparison
• Gpt 6 Sol Luna 50 Percent Price Cut Agentic Coding Cost Efficiency Frontier 2026 09 24 — GPT-6 family analysis
• Claude Opus 5 5 Fable 5 1 Performance 40 Percent Cost Reduction Pacing Frontier Safeguards 2026 09 23 — Claude Opus 5.5 cost reduction
• Deepseek V4 1 Flash Causal Encoder Decoder Kv Cache Compression Agentic Coding 2026 09 14 — DeepSeek V4.1-Flash architecture

External References

• BenchLM: MiniMax M3.1-Flash-Preview — Model tracking page
• OpenRouter — Model aggregation platform
• OpenCode — Open-source coding assistant

---

11. Future Directions

What to Watch

1. Public API release: The current restriction to Token Plan and MiniMax Code is temporary. A public API release with published pricing is expected.

2. Open-weight release: M3 is promised to be "fully open-sourced on HuggingFace and GitHub" soon. M3.1-Flash-Preview weights may follow.

3. Published benchmarks: The lack of published benchmarks is unusual for MiniMax (which published detailed benchmarks for M3). Expect official numbers in the coming weeks.

4. Competitive response: The stealth launch pattern may influence how other Chinese labs approach future releases.

5. Western lab adoption: Whether Western labs adopt the "ship first, document later" approach remains to be seen.

6. M3.1 full release: The "Preview" suffix suggests a full M3.1 release with published benchmarks and broader availability is coming.

7. Geopolitical impact: As Chinese models gain market share, expect increased regulatory scrutiny from the U.S. and EU.

8. Hardware diversification: The push toward domestic silicon (Huawei Ascend) may accelerate as Chinese labs scale inference.

---

Citation:

@misc{minimax_m3_1_flash_preview,
  title = {{MiniMax M3.1-Flash-Preview Launch}},
  url = {https://x.com/MiniMax_Agent},
  author = {{MiniMax}},
  month = {September},
  year = {2026}
}

@misc{minimax_m3,
  title = {{MiniMax M3: Coding \& Agentic Frontier}},
  url = {https://www.minimax.io/models/text/m3},
  author = {{MiniMax}},
  month = {June},
  year = {2026}
}

@software{minimax_code,
  title = {{MiniMax Code: Open-Source Coding Agent}},
  url = {https://github.com/MiniMax-AI/minimax-code},
  author = {{MiniMax}},
  month = {September},
  year = {2026}
}

---

This article was researched using official MiniMax sources (the M3 model page, platform documentation, GitHub releases, and X announcements) supplemented by independent testing reports from the developer community. All benchmark figures from independent testing are clearly attributed as such. The stealth benchmarking pattern and API accessibility findings are based on community-reported evidence.