Overview
The June 1, 2026 MiniMax release introduces M3 — what MiniMax describes as a frontier model that will become open-weight once the planned Hugging Face/GitHub release lands within roughly 10 days, bringing three capabilities together: top-tier coding and agentic performance, a 1M-token context window enabled by a new sparse attention architecture, and native multimodal training from step zero. Where M2.7 focused on recursive self-improvement and multi-agent collaboration at a 204K-token context, M3 represents a more fundamental architectural shift built around the MiniMax Sparse Attention (MSA) mechanism.
On the SWE-Bench Pro benchmark, MiniMax M3 reaches 59.0%, which the company reports surpasses GPT-5.5 and Gemini 3.1 Pro while approaching Claude Opus 4.7. In one widely cited internal demonstration, M3 autonomously reproduced an ICLR 2025 Outstanding Paper award winner over 12 hours of independent work and 18 commits — an end-to-end research engineering task on the AI agent frontier.
What's New
MiniMax Sparse Attention (MSA)
The headline architectural change in M3 is the new MSA mechanism, which partitions key-value blocks more precisely than prior sparse attention designs and uses a "KV outer gather Q" optimization for higher effective context coverage. The result is a 1M-token context window with per-token compute reduced to 1/20 that of the previous generation. Compared to M2.7's dense attention, M3 reports 9× speedup in prefilling and 15× speedup in decoding — making 1M-token workloads economically viable for the first time in MiniMax's lineup.
A guaranteed minimum of 512K tokens applies even under heaviest workloads, and the standard API rate covers requests up to 512K input tokens, with a separate long-context rate for the >512K range.
Frontier Coding & Agentic Performance
M3 posts the strongest software engineering benchmark numbers in MiniMax's history. On SWE-Bench Pro, M3 reaches 59.0%, which MiniMax reports as surpassing GPT-5.5 and Gemini 3.1 Pro and approaching Claude Opus 4.7. The MCP Atlas score of 74.2% positions M3 strongly on tool-using agent benchmarks.
Two flagship demonstrations highlight M3's autonomous agentic capabilities:
- ICLR 2025 paper reproduction: M3 independently reproduced an ICLR 2025 Outstanding Paper award winner over 12 hours of autonomous work, producing 18 commits across the full research engineering pipeline.
- 24-hour CUDA kernel optimization: M3 improved FP8 GEMM hardware utilization from 7.6% to 71.3% — a 9.4× speedup — through autonomous kernel iteration without human intervention.
Native Multimodality
Unlike retrofitted multimodal models, M3 was trained from step zero on a rebuilt data pipeline scaled to 100+ trillion tokens with interleaved multimodal content. The model supports image and video input alongside text, and adds computer use capabilities for agentic workflows that span screen interaction, document understanding, and visual reasoning.
On OmniDocBench, M3 scores above Gemini 3.1 Pro, and MiniMax reports that M3 surpasses Opus 4.7 on SVG-Bench — a structured visual generation benchmark that benefits coding workflows producing diagrams, UI mockups, and graphical assets.
Benchmark Performance
| Benchmark | M3 | Notes |
|---|---|---|
| SWE-Bench Pro | 59.0% | MiniMax-reported result; surpasses GPT-5.5 and Gemini 3.1 Pro in MiniMax's evaluation; approaches Opus 4.7 |
| Terminal-Bench 2.1 | 66.0% | Complex terminal task completion |
| SWE-fficiency | 34.8% | Software engineering efficiency measure |
| KernelBench Hard | 28.8% | Low-level GPU kernel generation |
| MCP Atlas | 74.2% | Model Context Protocol tool-use benchmark |
| OmniDocBench | Above Gemini 3.1 Pro | Multimodal document understanding |
| SVG-Bench | Surpasses Opus 4.7 | Structured SVG generation |
| Claw-Eval | Highest score | End-to-end autonomous agent evaluation |
Availability & Access
M3 is available through several channels on launch day:
| Access Path | Details |
|---|---|
| MiniMax Code | Coding agent / desktop-app experience designed for M3; its harness is based on OpenCode and Pi and is planned for future open-source release |
| MiniMax Agent | Free-tier access at agent.minimax.io with credit allocation |
| API | platform.minimax.io with standard (≤512K) and long-context (>512K) tiers |
| Token Plan | Subscription tiers — Plus / Max / Ultra |
| Open weights | Hugging Face release planned within ~10 days of launch (around 2026-06-11) |
The MiniMax Code harness is based on OpenCode and Pi, and MiniMax says it plans to open-source the project; full self-hosting will depend on the future weights, license, hardware requirements, and deployment documentation MiniMax publishes alongside the open-weight release.
Pricing & Plans
MiniMax M3 offers Token Plan subscriptions for predictable monthly usage alongside pay-as-you-go API access:
Token Plan subscriptions
| Plan | Monthly Price | Approx. Tokens |
|---|---|---|
| Plus | $20/month | ~1.7B tokens |
| Max | $50/month | ~5.1B tokens |
| Ultra | $120/month | ~9.8B tokens |
Pay-as-you-go API pricing
| Tier | Input ($/M) | Output ($/M) | Notes |
|---|---|---|---|
| Standard (≤512K input) | $0.60 | $2.40 | 7-day launch discount: $0.30 input / $1.20 output |
| Long-Context (>512K input) | $1.20 | $4.80 | Listed as limited / early-access on launch |
Cache-read pricing is listed separately on platform.minimax.io. Enterprise and high-volume teams should contact MiniMax directly for negotiated pricing.
Best For
- Research and engineering teams running long-context analysis workflows (whole-codebase reasoning, multi-document synthesis) where the 1M-token MSA architecture is the differentiator
- Open-source-friendly organizations evaluating frontier alternatives to closed-source models from OpenAI, Anthropic, and Google
- AI agent builders integrating autonomous coding, research, and computer-use workflows where M3's demonstrated 12-hour and 24-hour autonomous task runs are directly relevant
- Teams comparing MiniMax M3 against DeepSeek, Z.AI's GLM, and Kimi Claw on the open-weight Chinese LLM frontier
- Developers prototyping multimodal agent applications where native image, video, and computer use are part of the workflow rather than bolt-on capabilities
FAQ
How is M3 different from M2.7?
M2.7 focused on recursive self-improvement and native multi-agent collaboration at a 204,800-token context window. M3 is a more fundamental architectural shift: a new MiniMax Sparse Attention (MSA) mechanism expands the context window to 1M tokens, the data pipeline was rebuilt for native multimodal training from step zero, and per-token compute drops to roughly 1/20 of M2.7. On SWE-Bench Pro, M3's 59.0% advances beyond M2.7's 56.22% SWE-Pro result, though the two scores come from different benchmarks and should not be compared directly.
What is MiniMax Sparse Attention (MSA)?
MSA is a new sparse attention architecture introduced with M3. It partitions key-value blocks more precisely than prior sparse attention designs and uses a "KV outer gather Q" optimization, achieving higher effective context coverage. The practical impact is a 1M-token context window with per-token compute at 1/20 of the previous generation, plus 9× prefill speedup and 15× decode speedup. A guaranteed minimum of 512K tokens applies even under heavy workloads.
When will M3 weights be open-sourced?
MiniMax announced that M3 weights will be released on Hugging Face within approximately 10 days of the June 1, 2026 launch — placing the open-weight release around 2026-06-11. The technical report is planned for the same window. The MiniMax Code harness (built on OpenCode and Pi) is also planned for open-source release, though no date has been confirmed.
How does M3's coding performance compare to GPT-5.5 and Opus 4.7?
MiniMax reports M3 reaches 59.0% on SWE-Bench Pro, which the company says surpasses GPT-5.5 and Gemini 3.1 Pro and approaches Claude Opus 4.7. On SVG-Bench, M3 is reported to surpass Opus 4.7. These comparisons come from MiniMax-cited evaluations and warrant independent third-party validation. For teams running open-weight or self-hosted coding pipelines, M3 is positioned as the first model competitive with closed-source frontiers on these tasks.
Can I use M3 without an API key?
Yes. MiniMax Code and MiniMax Agent can be used through MiniMax's hosted product experience without wiring your own API key — MiniMax Agent at agent.minimax.io offers free-tier access with credit allocation. Programmatic integration still requires MiniMax Open Platform API access at platform.minimax.io, which provides standard (≤512K input tokens) and long-context (>512K input tokens) tiers.
