MiniMax M3 Brings Open-Weight Frontier AI with 1M Context

MiniMax released M3 on June 1, 2026, marking the first open-weight frontier model that combines 1M token context, native multimodal capabilities, and strong coding performance. It achieved 59.0% on SWE-Bench Pro, rivaling closed models like GPT-5.5. The API is already available, with weights planned for open release soon.

📑Table of Contents

1. Overview and Release Background of MiniMax M3
2. 1M Context and MSA Architecture Explained
3. Benchmark Comparison Table Including SWE-Bench Pro 59%
4. Native Multimodal and Agentic Capabilities
5. API Access, Pricing, and Limitations
6. Future Plans for the Open-Weight Version
7. Frequently Asked Questions (FAQ)

---

Overview and Release Background of MiniMax M3

MiniMax M3 was announced on the official blog on June 1, 2026. Unlike previous closed frontier models, it stands out as the first open-weight model offering frontier-level capabilities. Key features include 1M context length, native multimodal support for text/image/audio, and MSA (MiniMax Sparse Attention) for efficiency. It targets developers and agentic use cases specifically.

The release aligns with the broader industry trend toward open-source AI. MiniMax previously focused on closed APIs but chose to deliver frontier performance in an open-weight format with M3. This enables researchers and companies to customize the model in their own environments, offering advantages in privacy and cost control. Independent developer guides also highlight the significant impact of this open release on the AI community.

---

1M Context and MSA Architecture Explained

The standout technical feature of M3 is its 1M token context length, enabling analysis of long documents and large codebases. MSA optimizes the attention mechanism, delivering 9.7x faster prefill and 15.6x faster decode compared to the previous M2 model.

The architecture uses Fine-grained MoE with 229.9B total parameters, 9.8B active parameters, and 256 experts. This design maintains practical speeds even at 1M context while preserving high accuracy. The Lushbinary developer guide notes that this MoE approach contributes to the balance between cost efficiency and performance.

---

Benchmark Comparison Table Including SWE-Bench Pro 59%

Here are the key benchmark results for M3:

Benchmark	M3	GPT-5.5	Gemini 3.1 Pro	Source
SWE-Bench Pro	59.0%	58.6%	54.2%	Official (2026-06-01)
Terminal-Bench 2.1	66.0%	–	–	Official
SWE-fficiency	34.8%	–	–	Official
KernelBench Hard	28.8%	–	–	Official
MCP Atlas	74.2%	–	–	Official

Source: Official blog (2026-06-01), Lushbinary developer guide (2026-06-01)

M3 particularly excels in coding and agentic tasks, outperforming existing models and demonstrating strong real-world applicability.

---

Native Multimodal and Agentic Capabilities

M3 natively handles text, image, and audio as a multimodal model. It supports image recognition, audio input, video input, and even desktop computer operation. This allows a single model to manage diverse tasks efficiently.

Its agentic strengths are evident in high scores on Terminal-Bench and MCP Atlas, indicating strong performance in code execution and tool calling. Developers can leverage this for building automation scripts and agent workflows.

---

API Access, Pricing, and Limitations

The API is available at platform.minimax.io. Pricing is $0.60 / $2.40 per M tokens for up to 512K context, with a 50% off promotion in the first week.

Initial API support is limited to 512K context; full 1M support is planned after weights are released. Note that benchmark figures are vendor-provided, and independent verification after weights release is recommended. The Lushbinary guide also advises confirming pricing after the promotional period ends.

---

Future Plans for the Open-Weight Version

Weights are expected to be released on Hugging Face and similar platforms within about 10 days after API availability. The open release should accelerate fine-tuning and local deployments within the research community.

---

Frequently Asked Questions (FAQ)

Q1: On which benchmarks does MiniMax M3 surpass existing models?

It achieved 59.0% on SWE-Bench Pro, outperforming GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%).

Q2: Is the 1M context actually usable?

The official announcement claims 1M token support, but the initial API is limited to 512K, with full support planned after weights release.

Q3: What is the pricing?

$0.60 / $2.40 per M tokens (up to 512K), with a 50% off promotion in the first week.

Q4: When will the weights be opened?

The open-weight version is planned for release within about 10 days after API availability.

Q5: What advantages does it have over existing closed models?

It is the first open-weight model to combine frontier-level coding, long context, and multimodal capabilities, along with strong cost efficiency.

Related articles:

• Boltz Bio、BoltzMol-1 / BoltzProt-1 をリリース — 創薬向け新モデル、Claude Code / Codex 統合対応
• Claude Opus 4.8 リリース：Claude CodeのDynamic Workflowsと高速・低コスト化を解説
• Moonshot Kimi K2.7-Code リリース — 1Tパラメータオープンソース coding model が Claude Opus を上回る