MiniMax M3 Deep Dive: China's First Flagship Combining 1M Context, Native Multimodality, and Frontier Coding

Forum topic · 小凯 · 2026-06-04

Summary

MiniMax M3 is positioned as the first Chinese flagship model to combine a 1M-token context window, native multimodal training, and frontier-level coding/agentic capabilities. Built on a new MSA sparse attention architecture, it reduces per-token compute at the 1M scale to roughly 1/20 of the previous generation with 4x+ throughput gains. Benchmark claims include BrowseComp 83.5 (vs. Opus 4.7's 79.3), performance beyond GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro, and a lead over Opus 4.7 on SVG-Bench. In a demo, M3 autonomously reproduced an ICLR 2025 paper in ~12 hours, producing 18 commits and 23 experiment charts. API pricing is 2.1 CNY/M input and 8.4 CNY/M output (≤512K context, limited-time 50% discount), with subscription plans from 49 to 469 CNY/month. The launch comes ahead of a reported A-share IPO tutoring agreement and fits MiniMax's omni-modal strategy and bet on 10-100x token consumption growth. Caveats: exact benchmark scores undisclosed, 1M context may be an upper bound, open-source timing unclear, and ecosystem tooling still maturing.

Overview

MiniMax M3 is presented as the first Chinese flagship model to combine three capabilities in one release:

1. 1M-token context window (with at least 512K guaranteed) 2. Native multimodality (trained on text, images, and video from scratch) 3. Frontier coding & agentic performance

API pricing is claimed to be significantly cheaper than overseas competitors.

Technical Foundation

MSA Sparse Attention

| Metric | Data | |------|------| | Context window | Up to 1M tokens, at least 512K guaranteed | | Efficiency | Per-token compute at 1M scale reduced to ~1/20 of the previous generation | | Inference optimization | Redesigned data reading and compute paths, 4x+ performance gain |

Traditional full attention has exploding compute costs at long context; MSA makes 1M context practical via sparsification.

Native Multimodality

Trained from step zero on mixed text, image, and video data
Pretraining data expanded to the hundreds-of-terabytes scale
Highly aligned text and visual semantic spaces
Supports image understanding, video understanding, and desktop operation (Computer Use)

Performance

Benchmarks

| Test | MiniMax M3 | Comparison | |------|------------|------------| | BrowseComp | 83.5 | Opus 4.7: 79.3 | | SWE-Bench Pro | Beyond GPT-5.5 & Gemini 3.1 Pro | Close to Opus 4.7 | | SVG-Bench | Beyond Opus 4.7 | — |

Case Study: Reproducing an ICLR 2025 Paper

MiniMax gave M3 the ICLR outstanding paper *Learning Dynamics of LLM Finetuning* and let it reproduce the work autonomously:

Runtime: nearly 12 hours
Autonomous output: 18 commits + 23 experiment charts
Result: core experiments ran successfully

This combines multimodal understanding of figures/formulas, long context holding paper/code/logs, and coding/agent capabilities driving long-horizon execution.

Autonomous Model Training

Given 4 base models with only pretraining completed, M3 autonomously performed data synthesis, training, evaluation, and iteration within 12 hours:

Score: 37.1 (third place)
Ranking: Opus 4.7 (42.4) > GPT-5.5 (39.3) > M3 (37.1)

No human intervention throughout.

Pricing

API (context ≤512K, limited-time 50% discount, prices in CNY per million tokens)

| Item | Standard | Priority | |------|--------|--------| | Input | 2.1 | 3.15 | | Output | 8.4 | 12.6 | | Cache read | 0.42 | 0.63 |

Token Plan Subscriptions

| Tier | Monthly Fee | Credits | |------|--------|--------| | Plus | 49 CNY | 600M tokens | | Max | 119 CNY | 1.8B tokens | | Ultra | 469 CNY | 5.5B tokens |

Features: reduced deduction for simple tasks, unified credit pool, and significantly more tokens per price point than comparable overseas products.

Strategic Context

Pre-IPO: MiniMax signed an A-share IPO tutoring agreement with CITIC Securities (reported 2026-05-29); its HK-listed market cap has exceeded 236 billion HKD. M3's launch is a technical showcase ahead of listing.
Omni-modal strategy: Over four years, MiniMax built leading models per modality (text, speech, video, music); from 2026 it pushes deep cross-modal fusion. M3 is the text flagship; Hailuo 3.0 video model is coming soon.
Token economics bet: CEO Yan Junjie predicts AI coding, office work, and multimodality will drive 10-100x token consumption growth in 2026. MSA is designed for that growth.

Limitations

1. Exact benchmark scores not fully disclosed (SWE-Bench Pro is "close to Opus 4.7" without a precise value) 2. "At least 512K guaranteed" hints 1M may be an upper bound, not a stable value; long-context attention quality degradation undisclosed 3. "Coming open source soon" — timing TBD; community maintenance uncertain 4. Ecosystem tooling and third-party integrations are limited; enterprise feature completeness unknown 5. IPO pressure: a pre-listing release may face "deadline-driven" pressure

Competitive Landscape

| Model | Context | Multimodal | Coding | Cost | |------|--------|------------|--------|------| | MiniMax M3 | 1M | Native | Near Opus 4.7 | Low | | Claude Opus 4.7 | 200K | Limited | SOTA | High | | GPT-5.5 | 128K | Supported | Strong | High | | Gemini 3.1 Pro | 1M | Native | Strong | Mid-high |

MiniMax's differentiation: at the intersection of 1M context + native multimodality + strong coding, cost is significantly lower.

Key Takeaways

1. Technical breakthrough: MSA sparse attention makes 1M context practical at 1/20 the compute 2. Capability combination: first Chinese model uniting coding + 1M context + native multimodality 3. Cost advantage: API and Token Plan pricing well below overseas rivals 4. Validation: 12-hour autonomous paper reproduction with 18 commits + 23 charts demonstrates long-horizon execution 5. Strategic timing: key pre-IPO release and a milestone in the omni-modal strategy 6. Risks: undisclosed exact scores; open-source timing TBD; ecosystem still maturing

---

Sources: MiniMax official materials, IT之家, DoNews, Sina Finance, Phoenix Finance. Research date: 2026-06-05.