Release date: 2026-06-12 22:11 (Beijing time) Source: X post by MiniMax (@MiniMax_AI)
---
Event Overview
On the afternoon of June 12, 2026, MiniMax officially announced that MiniMax M3 open-weights model is now live on Hugging Face, described as "The First Open-Weights Model to Combine Three Frontier Capabilities."
- Model repository:
huggingface.co/MiniMaxAI/MiniMax-M3 - Technical paper:
huggingface.co/papers/2606.13392(titled "MiniMax Sparse Attention") - Engagement at posting: 379,000 views, 2,200 likes, 252 reposts
- Total parameters: ~428B
- Activated parameters: ~23B
- Activation ratio: ~5.4%, a typical sparse Mixture-of-Experts (MoE) architecture
- Classic Transformer attention scales as O(n²), making 1M tokens nearly infeasible
- Sparse Attention combines sliding windows, global anchors, and cross-layer sparsity to reduce attention computation toward O(n log n)
- Conceptually aligned with DeepSeek NSA, BGE HiAttention, and Mistral Sliding Window
- SWE-Bench Pro 59.0% — slightly above GPT-5 (~56%) and Claude Code (57%)
- Terminal Bench 2.1 66.0% — top tier among current open-source models
- MCP Atlas 74.2% — Anthropic's official MCP protocol evaluation; M3's score indicates strong native tool-call support, critical for the future agent ecosystem
- Training data scale and composition
- License type (Apache 2.0 or custom?)
- Commercial use restrictions
- API pricing (if any)
- 5.4% activation ratio keeps single-inference cost competitive
- 1M context enables true whole-repository comprehension rather than chunked processing
- MCP Atlas 74.2% allows M3 to plug directly into MCP tool ecosystems without adaptation overhead
- Early era (Llama 3): single capability + long context
- Mid era (DeepSeek V3): reasoning + coding
- Now (M3, Kimi-K2.7-Code): coding + agent + long-context as a unified trio
- DeepSeek NSA (Native Sparse Attention) deployed in V4
- MiniMax M3's proprietary Sparse Attention
- BGE, Mistral and others are also following
- IDE-based Coding Agents (Cursor, Cline, Roo Code): M3's Terminal Bench 2.1 66% combined with MCP Atlas 74% makes it an attractive foundation for in-IDE agents
- Cloud-based Coding Agents (Codex Cloud, Claude Code): open weights enable private deployment with data sovereignty, critical for regulated sectors like finance, government, and healthcare
- Chinese AI coding tool vendors: can build differentiated stacks such as "MiMo Code + M3 private deployment + custom prompts" without being locked to OpenAI or Anthropic
M3 is positioned as a "Coding & Agentic Frontier" model. MiniMax simultaneously published a set of agentic benchmark scores, aiming to become a foundational base model for AI programming and agent deployment.
---
Technical Deep Dive
1. Parameter Scale and Architecture
This scale places M3 in the top tier of open-source models—an order of magnitude larger than the 7B/8B-class MiMo Code models, and slightly lighter than flagship open models such as Llama 4 Maverick or DeepSeek V4, but with a tighter activation ratio (most MoE models operate at 8%–15% activation), meaning more controllable per-inference cost.
2. Core Innovation: MiniMax Sparse Attention
M3 uses the proprietary MiniMax Sparse Attention mechanism, which extends context length to 1M tokens. This is M3's key differentiator from comparable MoE models:
This provides structural compute advantages for long-document understanding, whole-repository code comprehension, and long-horizon agent tasks.
3. Official Agentic Benchmark Scores
| Benchmark | Score | Dimension Measured | |---|---|---| | SWE-Bench Pro | 59.0% | Real GitHub issue auto-fix | | Terminal Bench 2.1 | 66.0% | Terminal/CLI agent tasks | | SWE-fficiency | 34.8% | Software engineering efficiency (success × token economy) | | KernelBench Hard | 28.8% | GPU kernel code generation | | MCP Atlas | 74.2% | Model Context Protocol tool use |
Horizontal comparison reference:
4. Known Information Gaps
The official tweet did not disclose:
---
Why It Matters
1. Another "Extra-Large" Entry in the Chinese Open-Source Coding Race
In the first half of 2026, the Chinese open-source coding space is highly fragmented: MiMo Code, DeepSeek-Coder, Kimi-K2.7-Code, Qwen-Coder are all iterating rapidly. M3 enters with a 428B/23B MoE + Sparse Attention combination:
2. "Three-Capability Unity" Is the Real Watershed for the Agent Era
MiniMax positions M3 as the first open-weights model combining coding + agentic + long-context capabilities. The open-source evolution over the past year:
The bottleneck for Agentic AI in 2026 is no longer "can it code," but whether it can stably call tools, plan multi-step tasks, and understand long documents. M3's release timing matches the inflection point of agent deployment demand.
3. Sparse Attention Is the Next Technical Highland
Sparse attention mechanisms are evolving rapidly in large-model research:
If 2024's keyword was "MoE-ification," 2026's keyword is becoming "MoE + Sparse Attention"—extending MoE sparsity from the FFN layer to the attention layer, as a systematic engineering effort to reduce long-context inference cost. M3's choice to feature Sparse Attention as a core selling point reflects a bet on resolving the long-context-vs-cost tradeoff.
4. Impact on AI Coding Toolchains
In the short term, M3 will affect:
Risks and Open Questions
1. License not yet clarified: the open-source community's top concern; commercial usability depends on it 2. Training data compliance: 1M context + coding tasks require large amounts of GitHub data; compliance disclosure is critical 3. Inference cost: 428B/23B deployment needs H100/H200-class compute; per-token price determines practical usability 4. Synergy with other MiniMax products: will there be an M3-powered programming IDE or agent platform?
---
Summary
MiniMax M3 is not just another 400B MoE chasing leaderboards. Its differentiation lies in the "MoE + Sparse Attention + Three-Capability Unity" combination, targeting the core tension of the 2026 agent era: long context × tool calling × code execution. For anyone tracking the AI coding direction, this is one of the most noteworthy open-weights model releases of June.
If the license turns out to be permissive (Apache 2.0 / MIT class), M3 could become a de facto foundation model for Chinese AI coding toolchains.