⌘K
Change language Switch ThemeSign In
Narrow Mode
MiroMind Unveils SOTA Open-Source Agent Models MiroThinker-1.7 and H1 =====================================================================
MiroMind Unveils SOTA Open-Source Agent Models MiroThinker-1.7 and H1 =====================================================================  ### GitHubDaily
@GitHub_Daily
开源模型又诞生一匹黑马。MiroMind 团队发布最新研究型智能体模型：MiroThinker-1.7 和 H1。
在深度研究任务，以及专业科学与金融基准测试中全面 SOTA，超越 Claude 和 ChatGPT 等一众闭源模型。
GitHub：github.com/MiroMindAI/Mir…
模型下载：huggingface.co/collections/mi…
其中，1.7 模型在训练过程引入全新的「中期训练」阶段，极大提升了基础推理与规划的可靠性。
在此基础上，H1 模型进一步引入「重型推理能力」，让模型在推理和做计划时更加稳定，不再像以前的 AI 那样容易跑偏。
内置局部和全局的双重严苛验证机制，在思考过程中不断自我审视，过滤掉大量无效废话，让模型最终给出的建议变得极其可靠。
目前 1.7 模型权重已开源，并且即将在官网上提供体验，需要做深度调研的朋友，值得看下。
Web 端体验：dr.miromind.ai
APP 下载：miromind.ai/downloadShow More
Mar 17, 2026, 10:00 AM View on X
3 Replies
6 Retweets
19 Likes
2,357 Views  GitHubDaily @GitHub_Daily
One Sentence Summary
The MiroMind team has released research-oriented agent models MiroThinker-1.7 and H1, which comprehensively surpass closed-source models in in-depth research, scientific, and financial benchmarks, with partial weights already open-sourced.
Summary
This tweet introduces the MiroMind team's latest release of two research-oriented agent models: MiroThinker-1.7 and H1. These models have demonstrated outstanding performance in in-depth research tasks and professional scientific and financial benchmarks, achieving State-of-the-Art (SOTA) results, even surpassing leading closed-source models like Claude and ChatGPT. MiroThinker-1.7 incorporates a novel "mid-training phase," significantly enhancing the reliability of its fundamental reasoning and planning. Building upon this, the H1 model further integrates "robust reasoning capabilities," making it more stable in its inference and planning processes and less prone to deviation than previous AI models. Additionally, the models feature dual rigorous local and global validation mechanisms, enabling continuous self-scrutiny during the thought process to filter out irrelevant chatter and deliver exceptionally reliable recommendations. Currently, the model weights for MiroThinker-1.7 have been open-sourced, and web-based and app experience channels will soon be available, making them valuable resources for professionals engaged in deep research and AI development.
AI Score
88
Influence Score 12
Published At Today
Language
Chinese
Tags
MiroThinker
MiroMind
Open-Source Model
AI Agent
SOTA HomeArticlesPodcastsVideosTweets
MiroMind Unveils SOTA Open-Source Agent Models MiroThinke... ===============