MiniMax M3: When a Million-Token Scroll Meets a Coding Ghost — Triple Awakening
> On June 1, 2026, MiniMax delivered its answer. It is not just another bigger-parameter model, but a complete arsenal: a blade that can read a million lines of code, eyes that understand charts, and a soul that can work alone in a terminal for 24 hours.
---
📜 Part 1: What It Is
1.1 From abab to M3: Six Years in the Making
MiniMax didn't appear overnight. Founded in late 2021 in Shanghai by several ex-SenseTime engineers, the company's release cadence tells a story:
- January 2025: MiniMax-01 series open-sourced, introducing Lightning Attention — an attempt to push long-sequence complexity from quadratic to linear;
- June 2025: M1 released, 456B total / 45.9B activated parameters, 1M context — among the first Chinese models to make a 1M-token window standard;
- October 2025: M2 switched to Sparse MoE, activating only 10B of 230B parameters, cutting inference cost to ~8% of comparable closed-source models;
- February 2026: M2.5 became the flagship, hitting 80.2% on SWE-Bench Verified;
- March 2026: M2.7 began "recursive self-evolution" experiments, with the model participating in generating its own training data;
- June 1, 2026: M3 arrives — Coding + 1M context + native multimodality, all three, and open source.
- Read figures, formulas, and tables via multimodal capability;
- Remembered the entire paper + code + experiment logs in its 1M context, no chunking;
- Wrote runnable experiments from pseudocode;
- Iterated autonomously, confirming predicted probability trends matched the paper, observing the squeezing effect during the DPO phase, and validating the Extend mitigation.
- 147 benchmark submissions
- 1959 tool calls
- Autonomous progression: baseline implementation → autotune config generation → bottleneck diagnosis → CUDA Graph integration → persistent kernel rewrite → host-side scheduling optimization
The shrinking release intervals aren't rushed work — the self-evolution pipeline means models now help generate training data for other models, with humans only calibrating direction at key checkpoints.
> Tip: MoE (Mixture-of-Experts) is like a consulting firm with 230 experts, but each meeting only invites the 10 most relevant. Fewer people, smaller room, but decision quality holds — provided you invite the right people.
1.2 MSA: Sparse Attention Surgery
Transformer attention has a built-in flaw: computation grows quadratically with sequence length. Reading 1,000 tokens vs. 1,000,000 isn't a thousandfold difference — it's a millionfold. This \(O(N^2)\) curse is the first gatekeeper for all long-context models.
MiniMax's answer is MSA (MiniMax Sparse Attention). The core idea is simple: not every token needs to consult every other token. Filter coarsely first, then compute precisely.
MSA splits the KV cache into fixed-size blocks. Each query token first computes block-level affinity scores against block representatives; high-scoring blocks are kept, low ones discarded, and fine-grained attention runs only inside selected blocks.
It resembles Moonshot AI's MoBA but with different knife work:
| Dimension | MSA | MoBA | DSA | |------|-----|------|-----| | Routing granularity | Block-level coarse filter + full computation within blocks | Block-level top-k selection | Dynamic sparse patterns | | Hardware friendliness | High: contiguous KV block access, outer gather Q | Medium: scatter-gather overhead | Implementation-dependent | | Requires retraining | Yes | Yes | Yes | | Relation to Flash Attention | Fused at the operator layer | Grafted as preprocessing layer | Independent implementation |
A key design called "outer gather Q" aggregates matched queries with KV blocks as the outer loop. Each block is read only once with contiguous memory access; official claims say it's over 4x faster than Flash-Sparse-Attention and flash-moba.
> Tip: The KV cache is a Transformer's memory warehouse during inference. Every generated token produces a Key-Value pair stored for later reference. Longer context = bigger warehouse = slower lookups. Sparse attention essentially installs an indexing system.
1.3 Native Multimodality: Not a Bolted-On Puzzle
Many "multimodal models" are text models with vision modules grafted on afterward — like attaching a camera to a text-only typewriter.
M3 performs multimodal mixed training from step zero. Text, image, and video data alternate within the same sequence. MiniMax's experiments show interleaved data matters more than expected — text and images naturally interwoven, not artificially aligned.
The result: M3 can read a PDF paper (mixed text and figures), write code to reproduce experiments, and inspect output curves to adjust parameters — all through one neural pathway, not three plugins.
---
🔥 Part 2: What It Can Do
2.1 Coding: Engineering, Not Just Code
| Benchmark | Score | Meaning | |------|------|------| | SWE-Bench Pro | 59.0% | Real GitHub issue resolution rate | | Terminal Bench 2.1 | 66.0% | Terminal command-line agent task completion | | SWE-fficiency | 34.8% | Fixing issues with minimal changes | | KernelBench Hard | 28.8% | Geometric mean speedup of self-optimized CUDA kernels | | MCP Atlas | 74.2% | Tool-calling (MCP protocol) task completion |
SWE-Bench Pro at 59.0% exceeds GPT-5.5 and Gemini 3.1 Pro, approaching Claude Opus 4.7. But "approaching" needs unpacking: Opus 4.7's exact score isn't fully disclosed in public materials; extrapolating from GLM-5.1 comparison data puts it around 57–60%. Independent verification is still pending.
More notable are Terminal Bench 66.0% and MCP Atlas 74.2% — these measure not "writing code" but "using tools": running commands, calling APIs, checking docs, fixing bugs. That's the core of agent capability: a model that acts on environments, not just generates text.
2.2 Story One: 12 Hours, Paper to Reproduction
In internal testing, the most striking task was paper reproduction: replicating the ICLR 2025 Outstanding Paper *Learning Dynamics of LLM Finetuning*.
M3 ran autonomously for nearly 12 hours, producing 18 commits and 23 experiment charts, with no human intervention. It:
All four capabilities were indispensable: no multimodality means unreadable figures; no long context means the paper won't fit; no coding means nothing runs; no agent means nothing pushes the next round.
2.3 Story Two: 24 Hours, From Junk to Production-Grade CUDA Kernel
FP8 GEMM is the most compute-hungry operation in LLM inference. Hand-writing a production kernel on NVIDIA Hopper typically takes a senior team 1–2 weeks.
M3's starting conditions were brutal: only a task description, a benchmark script, and a non-functional Triton skeleton. No reference implementation to copy.
Within 24 hours:
Another detail: apart from Opus 4.7 and M3, most other models gave up within 30 submissions. M3's best solution appeared at submission 145, after multiple plateaus, still trying new directions. That stubbornness — knocking on doors in dead ends — is a hard metric for long-horizon autonomous iteration.
2.4 Story Three: M3 Training Models by Itself
If CUDA optimization is an exam with clear answers, model training is an open-book essay — no standard answer, only "try it."
PostTrainBench gave M3 four base models that had only completed pretraining, asking it to autonomously complete data synthesis → training → evaluation → iteration within 12 hours, across AIME2025 (math), BFCL (tool calling), GPQA (science), GSM8K (arithmetic), and HumanEval (code).
Final score: 0.37, slightly below Opus 4.7 (0.42) and GPT-5.5 (0.39), but clearly ahead of the rest. Not the strongest — but it walked the full path in open terrain where no answers were given.
2.5 Multimodal Results
M3 exceeds Opus 4.7 on SVG-Bench (SVG generation), Gemini 3.1 Pro on OmniDocBench (document understanding), and tops Claw-Eval (end-to-end agent evaluation). These numbers point to the same fact: M3's "eyes" and "hands" are matched — it reads design mockups and writes SVG code; reads chart-laden reports, extracts data, plots, and analyzes.
---
⚙️ Part 3: Pricing and Availability
3.1 MiniMax Code: Not Just a Chinese Claude Code
| Dimension | MiniMax Code | Claude Code | |------|-------------|-------------| | Core orchestration | Deep reflection and continuous error correction | Dynamic Workflows (JS-code fixed orchestration) | | Task decomposition | Agent Team multi-stage concurrent workflow | Single agent + sub-agents | | Multimodality | Native: Computer Use, can operate desktops | Mostly text/code | | Long-running | Producer + Verifier adversarial harness; can run for days | 5-hour window (Pro/Max) | | Open source | Harness planned to be open-sourced | Closed |
Claude Code is like a rigorous engineer; MiniMax Code is like a small squad: Producer writes code, Verifier picks bugs, they argue out a better version. Users can intervene anytime.
With M3's native multimodality, you could say on your phone: "open the local ERP client and batch-enter invoice data from this Excel" — and it completes the cross-app, cross-file operations on your computer. Computer Use as real dirty work, not a demo.
3.2 Token Plan: Cheap and Plentiful ≠ Better
| Tier | Price | Capacity | vs. Claude | |------|------|------|-----------| | Plus | ¥49/month | 600M tokens | ≈ 5x Claude Pro ($20) | | Max | ¥119/month | 1.8B tokens | ≈ 2x Claude Max 5x ($100) | | Ultra | ¥469/month | 5.5B tokens | ≈ 3x Claude Max 20x ($200) |
At equal prices, roughly 15x the usage of Claude subscriptions. But token price ≠ model quality. Thinking-mode latency, API stability, and reasoning consistency at long context matter too. The API supports thinking / non-thinking modes at the same pricing; a priority service tier (service_tier=priority) offers scheduling priority for SLA-sensitive industrial use (sales-gated now, full rollout within days).
3.3 API and Open Source: A 10-Day Promise
The M3 API is live, priced in two context-length tiers. MiniMax pledged to publish a technical report and open-source the model weights within 10 days.
That promise carries weight: M1, M2, M2.5, and M2.7 were all open-sourced on time under MIT license, verifiable on Hugging Face. Still, whether MSA will integrate smoothly with vLLM, SGLang, and other mainstream inference frameworks remains to be community-verified.
3.4 Limitations
Three at minimum:
First, benchmark context traps. Many M3 results were run on its own infrastructure using agent frameworks like Claude Code, Mini-SWE-Agent, and Terminus — mixing "model capability" with "scaffolding capability." Scores may shift with a different framework. Until independent replication, treat numbers as reference, not verdicts.
Second, attention quality at long context. MiniMax says "the vast majority of capabilities match full attention" at 1M — but "vast majority" leaves a back door. No disclosed degradation curves for MSA's needle-in-a-haystack precision at extreme lengths.
Third, "seeing" ≠ "understanding." M3 can read figures in papers, but can it grasp the statistical assumptions behind them, or distinguish significance markers from noise? Native multimodality solves format compatibility; semantic depth takes longer to prove.
---
🌌 Epilogue
MiniMax M3 isn't about bigger parameters — it's a capability-combination experiment: Coding + 1M context + native multimodality, all three, open source. Overseas closed frontier models have all three routinely; among Chinese models, M3 is the first to combine them and open up. This isn't a "surpass GPT" narrative — it's a "completing the puzzle" narrative.
The three case studies — paper reproduction, CUDA optimization, autonomous training — reveal a trend: next-gen coding agent competition lies not in single-shot code accuracy, but in long-thread collaboration, planning, and human-agent synergy. Working alone for 12 or 24 hours, persisting through plateaus, making autonomous decisions on open problems — that's the qualitative shift from "tool" to "partner."
If the 10-day open-source promise holds, the community will, for the first time, run a model with coding, 1M context, and multimodality on local GPUs. That's not consumption — that's takeover.
> "Model updates today move so fast it's easy to forget this remains a steady, brick-by-brick endeavor. It has its own objective laws, and it rewards teams that advance solidly along them." > > — From the closing of MiniMax's M3 release blog post
---
📚 References
1. MiniMax M3 Official Release Blog (2026-06-01). https://www.minimaxi.com/blog/minimax-m3 2. Liu et al. (2025). *MoBA: Mixture of Block Attention for Long-Context LLMs*. Moonshot AI. 3. DeepSeek-AI (2025). *Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention*. 4. Jimenez et al. (2024). *SWE-bench: Can Language Models Resolve Real-World GitHub Issues?* ICLR 2024. 5. Ouyang et al. (2025). *KernelBench: A Benchmark for Automatic Kernel Optimization*.