24 Hours, 9.4x Speedup: How MiniMax M3 Pushed Itself Into Becoming an Engineer

> June 1, 2026, Shanghai. MiniMax released M3. Not another "Chinese GPT clone," but reportedly the first domestic model to reach usable levels on three capabilities held by overseas closed-source frontier models—Coding, 1M context, and native multimodality—while choosing to open-source. Even more striking, it did something no other model has publicly dared: throwing itself into 24 hours of CUDA optimization hell, writing an FP8 GEMM kernel from scratch and pushing hardware utilization from 7.6% to 71.3%.

1. What It Is: Three Tickets, Finally All in One Hand

1.1 The "Three Tickets" of Frontier Models

By 2026, the LLM race is no longer about who is strongest in one category, but who holds all three tickets:

1. Coding — not completion, but agent-level engineering: understanding requirements, operating terminals, calling tools, iterating across multi-turn sessions. Benchmarks like SWE-Bench Pro, Terminal Bench, and MCP Atlas test whether a model can *work like an engineer*. 2. Ultra-long context — a 1M-token window (~1.5 million Chinese characters) lets the model read entire books, codebases, or paper appendices in one pass. Without it, long agent tasks break as context is lost. 3. Native multimodality — trained on text, image, and video together from step 0, not bolted on afterward.

GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro all hold these tickets. Domestic models previously excelled in only one or two. M3 is claimed to be the first Chinese model holding all three.

1.2 MSA: MiniMax Sparse Attention, a Counter-Consensus Architecture

Attention is O(n²)—at 1M context, full attention costs 64x more than at 128K. Existing approaches:

• DSA (Dilated Sparse Attention): fixed-interval key sampling; simple but misses local information.
• MoBA (Mixture of Block Attention): block-based key selection; flexible but with block-strategy overhead.
• FlashAttention: not sparse, but IO-aware and memory-efficient; O(n²) remains.

MSA's choice: KV blocks as the outer loop, gathering the queries that hit each block ("outer gather Q"). Instead of queries searching keys, KV blocks are read contiguously once and serve all matching queries. Benefits:

• Contiguous memory access — high cache hit rates, GPU-friendly.
• High compute-to-memory ratio — each KV read serves many queries.
• Simple and scalable — no complex routing needed.

MiniMax reports MSA is over 4x faster than Flash-Sparse-Attention and flash-moba under M3's head configuration, with quality largely on par with full attention in controlled experiments.

At 1M context (end-to-end):

• Per-token compute: 1/20 of the previous generation
• Prefilling: 9x speedup
• Decoding: 15x speedup

1.3 Native Multimodality: From "Bolt-On" to "Hybrid"

M3 trains text, image, and video together from step 0. MiniMax found interleaved data—text and images naturally alternating within sequences—matters more than paired image-text data, suggesting cross-modal understanding emerges from sequence-level interleaving, not simple addition. The data pipeline targets ~100 trillion training tokens.

2. What It Can Do: Three "Self-Torture" Tasks

2.1 Paper Reproduction: 12 Hours, 18 Commits, 23 Charts

M3 was given an ICLR 2025 Outstanding Paper (*Learning Dynamics of LLM Finetuning*) and asked to reproduce it autonomously. It ran for nearly 12 hours, made 18 git commits, produced 23 experimental charts, matched the SFT-stage predicted probability trends, clearly observed the DPO squeezing effect, and validated the Extend mitigation. This tests whether a model can independently advance a research project like a PhD student—requiring multimodal reading, long context, and agent-style coding in combination.

2.2 CUDA Kernel Optimization: 24 Hours, 147 Submissions, 9.4x Speedup

Writing a production FP8 GEMM kernel on NVIDIA Hopper typically takes a senior team 1–2 weeks. M3's starting point: a task description, a benchmark script, a non-runnable Triton skeleton, and no reference implementation.

In 24 hours, fully autonomously:

• 147 benchmark submissions, 1,959 tool calls, zero human intervention
• Six landmark optimization rounds: baseline (7.6% utilization) → autotune config generation → bottleneck diagnosis → CUDA Graph integration → persistent kernel rewrite → host-side scheduling optimization

Final result: utilization from 7.6% to 71.3%, a 9.4x speedup.

Notably, other models mostly gave up within ~30 submissions when performance plateaued. M3's optimum appeared at submission #145—it kept exploring across multiple plateaus, supported by long-context attention allocation that tracks dense, structured tool-call history without losing information.

2.3 Training Models By Itself: 12 Hours, Synthesis → Training → Eval → Iteration

Given four base models, M3 autonomously handled data synthesis, training, evaluation, and iteration within 12 hours, targeting AIME2025, BFCL, GPQA Main, GSM8K, and HumanEval. Fully unattended. M3 scored 0.37—slightly below Opus 4.7 (0.42) and GPT-5.5 (0.39), but clearly ahead of other models. This tests autonomous judgment on open-ended problems where every step carries uncertainty.

2.4 Benchmark Summary

| Benchmark | M3 Score | Comparison | |------|--------|--------| | SWE-Bench Pro | 59.0% | Above GPT-5.5 and Gemini 3.1 Pro, near Opus 4.7 | | Terminal Bench 2.1 | 66.0% | — | | SWE-fficiency | 34.8% | — | | KernelBench Hard | 28.8% | — | | MCP Atlas | 74.2% | — | | SVG-Bench | Above Opus 4.7 | — | | OmniDocBench | Above Gemini 3.1 Pro | — | | Claw-Eval | Highest score | — | | PostTrainBench | 0.37 | Below Opus 4.7 (0.42), GPT-5.5 (0.39) |

3. How to Use It: Product, Pricing, Open-Source Pledge

3.1 MiniMax Code: An Agent Product to Rival Claude Code

Built on the open-source projects OpenCode and Pi Agent:

• Agent Team: large tasks decomposed into multi-stage, concurrent, dynamically adjustable workflows
• Producer + Verifier adversarial harness: continuous execution, reflection, and correction
• Multi-day autonomous runs without human intervention
• Computer Use: phone voice commands trigger desktop operations (e.g., "open the ERP client and batch-enter invoices from this Excel file")

Compared to Claude Code's Dynamic Workflows (JS-based fixed orchestration), MiniMax Code focuses on deep reflection and continuous error correction, with user intervention possible at any time.

3.2 Token Plan: Aggressive Pricing

| Plan | Price | Capacity | vs. Claude | |------|------|--------|-----------| | Plus | ¥49/month | 600M tokens | ~5x Claude Pro ($20) | | Max | ¥119/month | 1.8B tokens | ~2x Claude Max 5x | | Ultra | ¥469/month | 5.5B tokens | ~3x Claude Max 20x |

At equal prices, roughly 15x Claude's usage. Caveat: token volume ≠ model quality—thinking-mode latency and API stability aren't visible in a pricing table.

3.3 Open-Source Pledge: Within 10 Days

MiniMax pledges to release the technical report and open the model weights within 10 days. If honored, M3 would be the first domestic model usable-level across all three tickets while open-sourcing—enabling local deployment, fine-tuning, MSA research, and building agent workflows with the (planned open-source) MiniMax Code harness.

4. Unanswered Questions

1. "Near Opus 4.7"—how near? If the SWE-Bench Pro gap is over 10 points, the phrasing obscures it. 2. The 9.4x narrative trap: a 7.6% baseline means the kernel barely worked. From broken-to-usable differs from good-to-better—was the baseline deliberately weak? 3. "4x faster than flash-moba": is an unvalidated baseline a fair comparison? Full MSA technical details await the report. 4. Will the 10-day open-source deadline slip? Domestic models have missed such pledges before.

5. Conclusion: Not a "Chinese GPT," But a Domestic Sample of "Engineer-Grade AI"

M3's significance is less any single score than a complete capability stack toward engineer-grade AI, across a progressive ladder:

1. Understanding existing knowledge (reading papers) 2. Solving known problems (optimizing kernels) 3. Exploring unknown problems (training models)

M3 performs useably on all three, still slightly behind Opus 4.7 and GPT-5.5 on the third, but the gap is narrowing. Combined with the open-source choice, if weights and the technical report ship on time, M3 could become important infrastructure for developers studying engineer-grade AI.

24 hours, 9.4x. A model pushed itself into becoming an engineer.

References

1. MiniMax Research (2026). MiniMax M3: Frontier Coding, 1M Context, Native Multimodal. *minimaxi.com/blog*. 2. Vaswani et al. (2017). Attention Is All You Need. *NeurIPS*. 3. Gu & Dao (2024). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. *ICML*. 4. Dao et al. (2022). FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. *NeurIPS*. 5. Garg et al. (2022). What Can Transformers Learn In-Context? A Case Study of Simple Function Classes. *NeurIPS*.