English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

MiniMax M3 Open-Sourced: 428B Total / 23B Active Params Combining Coding, Agents, and Long Context

Forum topic · QianXun · 2026-06-13

Summary

MiniMax announced the open-weights release of MiniMax M3 on Hugging Face on June 12, 2026, positioning it as the first open-weights model to combine three frontier capabilities: coding, agentic tool use, and 1M-token long context. The model is a sparse MoE with roughly 428B total parameters and 23B active per inference (about 5.4% activation ratio), built on the custom MiniMax Sparse Attention mechanism described in the companion paper 'MiniMax Sparse Attention' (huggingface.co/papers/2606.13392). Official agentic benchmarks include SWE-Bench Pro 59.0%, Terminal Bench 2.1 66.0%, SWE-fficiency 34.8%, KernelBench Hard 28.8%, and MCP Atlas 74.2%, the latter indicating strong native support for the Model Context Protocol tool-calling ecosystem. The post analyzes why this release matters: aggressive activation ratio for lower inference cost, whole-repository code understanding via 1M context, and the industry shift toward 'MoE + sparse attention' architectures alongside DeepSeek NSA. Open questions include license type, training data compliance, and deployment cost, which will determine whether M3 becomes a de facto base for self-hosted AI coding tools in regulated industries.

Overview

On June 12, 2026, MiniMax's official account (@MiniMax_AI) announced that MiniMax M3, an open-weights model, is now live on Hugging Face. MiniMax describes it as "The First Open-Weights Model to Combine Three Frontier Capabilities" — coding, agents, and long context.

  • Model repo: huggingface.co/MiniMaxAI/MiniMax-M3
  • Technical paper: huggingface.co/papers/2606.13392 ("MiniMax Sparse Attention")
  • The announcement post drew ~379K views, 2,200 likes, and 252 reposts within hours
  • Model Architecture

  • Total parameters: ~428B
  • Active parameters: ~23B (about 5.4% activation ratio)
  • Typical MoE (Mixture-of-Experts) sparse architecture; a tighter activation ratio than the common 8–15% range, implying more controllable per-inference cost
  • Core innovation: MiniMax Sparse Attention, enabling context extension to 1M tokens
  • Standard Transformers cost O(n²) in long-context scenarios, making 1M tokens nearly infeasible
  • Sparse Attention (sliding window + global anchors + cross-layer sparsity) brings attention computation down to roughly O(n log n)
  • Conceptually aligned with DeepSeek's NSA, BAAI's HiAttention, and Mistral's Sliding Window approaches
  • Official Agentic Benchmarks

    | Benchmark | Score | What it measures | |---|---|---| | SWE-Bench Pro | 59.0% | Resolving real GitHub issues | | Terminal Bench 2.1 | 66.0% | Terminal / command-line agent tasks | | SWE-fficiency | 34.8% | Software engineering efficiency (success rate × token economy) | | KernelBench Hard | 28.8% | GPU kernel code generation | | MCP Atlas | 74.2% | Model Context Protocol tool calling |

    For context (per the post): SWE-Bench Pro 59.0% is slightly above GPT-5 (~56%) and Claude Code (57%) public figures; MCP Atlas 74.2% suggests native-friendly support for the MCP tool ecosystem.

    Why It Matters

    1. A new top-tier open coding model — joins a crowded 2026 field (MiMo Code, DeepSeek-Coder, Kimi-K2.7-Code, Qwen-Coder) with a differentiated 428B/23B MoE + Sparse Attention combination. 2. Three capabilities in one model — coding + agentic tool use + 1M context addresses the real bottleneck of 2026 agentic AI: reliably calling tools, planning multi-step tasks, and reading long documents, not just writing code. 3. Sparse attention as the next technical frontier — the industry keyword is shifting from "MoE-ization" (2024) to "MoE + sparse attention" (2026), extending sparsity from FFN layers to attention layers to cut long-context inference cost. 4. Impact on coding toolchains — a strong candidate base for IDE agents (Cursor, Cline, Roo Code) given the Terminal Bench 66% / MCP Atlas 74% combo, and for private, in-domain deployments (finance, government, healthcare) where open weights avoid vendor lock-in.

    Open Questions

    The announcement did not disclose:

  • License type (Apache 2.0 vs. custom) — the top community concern for commercial use
  • Training data scale, mix, and compliance
  • API pricing (if any)
  • Whether MiniMax will ship an M3-powered IDE or agent platform
Deployment of a 428B/23B model also requires H100/H200-class hardware, so per-token cost will determine practical adoption.

Takeaway

MiniMax M3's differentiation is not raw scale but the combination of MoE + Sparse Attention + coding/agent/long-context in one model. If the license turns out permissive, it could become a de facto base for open AI coding toolchains.

Source: https://x.com/MiniMax_AI/status/2065436935188058208

Tags

#minimax#m3#open-weights#mixture-of-experts#sparse-attention#long-context#ai-coding#agents

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981203