MiniMax M3.1 Flash: 428B Open MoE, Video Input & Benchmarks

Executive Summary

An architectural deep dive into MiniMax M3.1 Flash, covering MiniMax Sparse Attention across 1M tokens, MXFP8 and NVFP4 quantization-aware training, vLLM and NVIDIA NIM serving, and DeepSWE, BankerToolBench, and CyberGym benchmarks.

From "Space Bunny" to Open Frontier: What Is MiniMax M3.1 Flash?

MiniMax M3.1 Flash is a native multimodal Mixture-of-Experts (MoE) model formally released by MiniMax on October 19, 2026, with open weights published on Hugging Face and ModelScope shortly afterward. Before unmasking, the checkpoint ran as the viral stealth preview "Space Bunny", processing over 200 trillion tokens across live coding and agentic harnesses. Architecturally, MiniMax M3.1 Flash pairs 428B total parameters with 23B active parameters per token (5.37% activation), natively ingesting text, multi-image sets, and continuous video inside a 1,000,000-token context window with 128,000 max completion tokens. Powered by MiniMax Sparse Attention and quantization-aware training (QAT) in MXFP8 and NVFP4, the architecture includes day-one vLLM and NVIDIA NIM support plus a tunable reasoning effort ladder (low to max, defaulting to max). On frontier evaluations, MiniMax M3.1 Flash hits 43 on the Artificial Analysis Intelligence Index ($0.22 task cost), 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 76.2 on BankerToolBench, and ~82% on CyberGym, priced at $0.15/M input, $0.60/M output, and $0.015/M cached input (70% cheaper than prior MiniMax generations).

Key Takeaways

Unmasked After 200+ Trillion Stealth Tokens as "Space Bunny"

Before its October 19, 2026 release and open-weight upload to Hugging Face and ModelScope, MiniMax M3.1 Flash proved real-world agentic reliability as the viral "Space Bunny" preview.

428B Total / 23B Active Native Video Multimodal MoE

Activating 23B of 428B parameters per token, the sparse backbone ingests text, multi-image arrays, and native video across a 1,000,000-token window with 128,000 max output tokens.

MiniMax Sparse Attention + MXFP8 and NVFP4 QAT

MiniMax Sparse Attention cuts 1M-context KV bandwidth, while MXFP8 and NVFP4 quantization-aware training enables lossless low-bit serving via vLLM and NVIDIA NIM.

73.7% DeepSWE v1.1 and 43 Artificial Analysis Index

Defaulting its reasoning effort ladder to max, MiniMax M3.1 Flash scores 43 on Artificial Analysis ($0.22 task cost) and resolves 73.7% of DeepSWE v1.1 software issues.

76.2 BankerToolBench, ~82% CyberGym & 61.7 OSWorld V2

Across enterprise domains, the model posts 76.2 on BankerToolBench, ~82% on CyberGym security patching, and 61.7 on OSWorld V2 desktop control.

Flat $0.15/M Input & $0.015/M Cache (~70% Generational Cut)

Commercial endpoints for MiniMax M3.1 Flash charge $0.15/M input, $0.60/M output, and $0.015/M cached reads (~70% below prior MiniMax models) with zero >100K cliff.

Architectural & Engineering Deep Dive

428B/23B Native Multimodal MoE and the 200T-Token "Space Bunny" Crucible

Rather than bolting a frozen vision encoder onto a text model late in post-training, MiniMax M3.1 Flash was trained natively across interleaved text, multi-image sets, and temporal video tokens inside a 428B-parameter Mixture-of-Experts transformer that activates just 23B parameters per token (~5.37%). This high total-to-active ratio gives MiniMax M3.1 Flash deep domain memory across obscure SDKs, CVEs, and banking rules while keeping per-token FLOPs at a compact 23B budget. Prior to its October 19, 2026 unveiling and open-weight release on Hugging Face and ModelScope, the checkpoint processed over 200 trillion live tokens as "Space Bunny", sharpening RL tool recovery and its tunable reasoning effort ladder (low to max, defaulting to max for 43 AA Index reasoning).

MiniMax Sparse Attention Across 1,000,000 Multimodal Tokens

Combining continuous video timelines with 500,000-token monorepos quickly exhausts dense quadratic attention and HBM KV caches. MiniMax M3.1 Flash overcomes this via MiniMax Sparse Attention, pairing local sliding-window fidelity with learned block-sparse routing over long-range semantic anchors. Across prefill and 128,000-token decoding, attention heads bypass redundant background video patches and boilerplate syntax while retaining exact cross-modal recall across 1,000,000 tokens. Near-linear long-context scaling enables MiniMax M3.1 Flash to sustain flat API rates—$0.15/M input, $0.60/M output, and $0.015/M cached reads (~70% below prior generations)—without long-context pricing cliffs.

Quantization-Aware MXFP8 & NVFP4 Training for vLLM and NVIDIA NIM

Hosting 428B BF16 weights normally demands >850 GB of VRAM before KV cache, while naive post-training INT4 quantization degrades expert router gating on multi-step coding tasks. MiniMax M3.1 Flash eliminates this trade-off through quantization-aware training (QAT) targeting microscaling MXFP8 and NVIDIA Blackwell NVFP4 formats during pre-training and RL. Consequently, open MXFP8 and NVFP4 checkpoints retain full benchmark parity on DeepSWE v1.1 (73.7%), BankerToolBench (76.2), and CyberGym (~82%), deploying out of the box on 4x B200 or H200 nodes via day-one vLLM and NVIDIA NIM support.

Comprehensive Model Specifications

Head-to-Head Performance & Benchmark Matrix

Real-World Implementation & Hands-on Verification

Native Video + Repository Bug Reproduction from QA Screen Recordings

Scenario Evaluation: A frontend reliability team feeds a 4-minute 60fps Playwright failure video and 310,000 tokens of Next.js source code directly into MiniMax M3.1 Flash at default max effort.

Standardized Benchmark Prompt:

Analyze the checkout screen recording, correlate the visual race condition at 02:14 with Apollo cache normalization in src/checkout/, and emit a regression-tested TypeScript patch.

Empirical Output Summary: Across one 385,000-token multimodal pass, MiniMax M3.1 Flash spotted a 180ms stale optimistic UI write across three frames, fixed a missing keyFields entry, and wrote a 22-line patch.

Evaluation Verdict: Native video + 73.7% DeepSWE coding in MiniMax M3.1 Flash removed transcription middleware and cost $0.012 on cached follow-ups.

Air-Gapped Vulnerability Patching on NVIDIA NIM (NVFP4)

Scenario Evaluation: An AppSec team hosts the open NVFP4 checkpoint of MiniMax M3.1 Flash on an 8x B200 Blackwell node via NVIDIA NIM and vLLM to audit C/Rust network daemons offline.

Standardized Benchmark Prompt:

Audit the 420,000-token packet parser repo and fuzzing crashes, construct a minimal PoC reproducing the heap over-read, and synthesize a bounds-checked patch.

Empirical Output Summary: Activating 23B parameters under MiniMax Sparse Attention, MiniMax M3.1 Flash isolated an integer wrap-around in TLV length parsing, verified crash reproduction, and passed all sanitizers.

Evaluation Verdict: Validating its ~82% CyberGym score, NVFP4 QAT preserved bit-exact arithmetic inside an air-gapped enclave.

Institutional Treasury Orchestration on BankerToolBench Schemas

Scenario Evaluation: A quantitative treasury desk evaluates MiniMax M3.1 Flash across 45 chained REST and SQL tools spanning FX quotes, Basel III liquidity ratios, and settlement ledgers.

Standardized Benchmark Prompt:

Reconcile intraday EUR/USD margin calls across four broker APIs, verify counterparty caps against policy tables, and stage idempotent settlement payloads.

Empirical Output Summary: Over 29 tool calls, MiniMax M3.1 Flash preserved strict JSON schemas, caught a duplicate settlement nonce, and dropped to low effort on formatting turns to cut latency 38%.

Evaluation Verdict: Confirming its 76.2 BankerToolBench rating, the pipeline paired financial rigor with $0.015/M cached prefixes.

Evidence Ledger: Fact-Check & Verification Audit

To ensure search engine E-E-A-T integrity, claims are classified across confirmed, reported, unverified, and unknown tiers:

Production Caveats & Known Constraints

428B Total Weights Require Data-Center Multi-GPU Nodes

Even in NVFP4 (~220–240 GB VRAM before KV cache), self-hosting all 428B expert parameters of MiniMax M3.1 Flash requires multi-GPU server nodes rather than single desktop cards.

Default Max Reasoning Effort Adds Latency on Trivial Prompts

Because MiniMax M3.1 Flash defaults reasoning effort to max, omitting explicit low effort on simple routing or tagging calls incurs unnecessary thinking tokens.

OSWorld V2 (61.7) Trails Specialized Proprietary GUI Endpoints

While 61.7 on OSWorld V2 leads open multimodal peers, ultra-dense pixel RPA tasks may still see higher click precision from specialized closed GUI models than MiniMax M3.1 Flash.

Recommended Developer Action Plan

Deploy NVFP4 or MXFP8 Weights with vLLM or NVIDIA NIM

Pull official quantized checkpoints of MiniMax M3.1 Flash from Hugging Face or ModelScope onto Hopper or Blackwell nodes for private 1M-context serving.

Modulate Reasoning Effort Between Low and Max Per Route

Retain default max effort on MiniMax M3.1 Flash for DeepSWE refactoring and CyberGym security audits, while setting low effort on latency-critical classifiers.

Unify Video + Repository Debugging Under $0.015/M Prefix Caching

Stream MP4 bug captures and repository trees directly into MiniMax M3.1 Flash with prompt caching enabled to replace fragile multi-stage vision-to-code pipelines.

Frequently Asked Questions

What is MiniMax M3.1 Flash and what was its "Space Bunny" codename?

MiniMax M3.1 Flash is MiniMax's 428B-parameter (23B active) open-weight multimodal MoE model officially released on October 19, 2026. Before launch, it ran in viral stealth preview as "Space Bunny", processing over 200 trillion tokens.

Are the model weights for MiniMax M3.1 Flash openly available for self-hosting?

Yes. MiniMax published open weights for MiniMax M3.1 Flash on both Hugging Face and ModelScope shortly after its October 19, 2026 release, including quantization-aware MXFP8 and NVFP4 checkpoints with day-one vLLM and NVIDIA NIM support.

How does MiniMax M3.1 Flash perform on DeepSWE v1.1, CyberGym, and BankerToolBench?

MiniMax M3.1 Flash achieves 43 on the Artificial Analysis Intelligence Index ($0.22 task cost), 73.7% on DeepSWE v1.1, 76.2 on BankerToolBench, ~82% on CyberGym, and 61.7 on OSWorld V2.

Does MiniMax M3.1 Flash support native video input alongside text and images?

Yes. MiniMax M3.1 Flash is a native multimodal MoE accepting text, multi-image sets, and continuous video input across a 1,000,000-token context window, generating up to 128,000 text output tokens.

What is the API pricing for MiniMax M3.1 Flash?

Hosted API access for MiniMax M3.1 Flash costs $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens—approximately 70% cheaper than previous MiniMax generations.

How do MiniMax Sparse Attention and MXFP8/NVFP4 quantization improve inference?

MiniMax Sparse Attention slashes KV-cache memory bandwidth across million-token sequences, while quantization-aware training in MXFP8 and NVFP4 preserves full reasoning accuracy on 4-bit Blackwell and 8-bit Hopper GPUs.

How does the tunable reasoning effort ladder work in MiniMax M3.1 Flash?

MiniMax M3.1 Flash exposes an API reasoning effort setting ranging from low to max, defaulting to max for highest coding and agentic accuracy while allowing low effort to minimize latency and output token usage.

What hardware is recommended to self-host MiniMax M3.1 Flash in production?

Using vLLM or NVIDIA NIM, the NVFP4 checkpoint of MiniMax M3.1 Flash fits comfortably on 4x Blackwell B200 or 4x H200 GPUs (~220–240 GB weight footprint plus KV cache), while MXFP8 targets 8x H100/H200 nodes.

Verified Sources & References

1. [MiniMax Official Announcements] Introducing MiniMax M3.1 Flash: Unmasking Space Bunny with 428B Native Multimodal MoE (ID: src-minimax-m31-rel)
2. [Hugging Face Model Hub] MiniMaxAI / MiniMax-M3.1-Flash Open Weights, MXFP8 & NVFP4 Model Card (ID: src-minimax-m31-hf)
3. [ModelScope Open Model Repository] MiniMax / MiniMax-M3.1-Flash Official Checkpoints & Deployment Guide (ID: src-minimax-m31-ms)
4. [Artificial Analysis] MiniMax M3.1 Flash Intelligence Index, Speed, and $0.22 Task Cost Benchmark (ID: src-aa-minimax-m31)
5. [MiniMax Developer Platform] MiniMax Commercial API Pricing, 1M Context & $0.015/M Prompt Caching Rates (ID: src-minimax-api-pricing)
6. [NVIDIA Developer & vLLM Ecosystem] Deploying MiniMax M3.1 Flash with NVFP4, MXFP8, vLLM, and NVIDIA NIM (ID: src-nvidia-nim-minimax)