📰 Minerve Blog

Version 1.11 introduces a brand-new tool discovery system inspired by Anthropic's approach, alongside major improvements to reasoning, stability, and the overall user experience.

🔍 Tool Discovery System (Preview)

• New tool search system now available in preview, inspired by Anthropic's model of tool discovery.
• Core tools remain always accessible while Minerve can now discover specialized tools on demand — enabling near-infinite tool capacity without overhead.
• Tested with 170 tools and 650 functions. Working smoothly.
• Minerve Preview now equipped with a search-and-execute tool to dynamically locate and invoke specialized tools.

🧠 Reasoning & Prompt System

• Full reasoning chain awareness — Minerve now has access to its entire chain of prior reasoning, not just the last step. Significant gains on multi-step tasks and long-context workflows.
• Minerve Preview & Preview Heavy system prompt improved with new directives:
- Read or create README files automatically. - Verify Python file compilation after creation or modification. - Reserve the internal agent only for tasks requiring multiple tool calls. - Never delegate destructive or critical operations that require full context. - Integrated new coding directives inspired by Andrej Karpathy. - Improved file search directives (*filename patterns).

🛡️ Stability & Safety

• Shell access lock — prevents accidental deactivation of shell access.
• Infinite tool-call loop protection — if the model repeats the same tool call more than twice, a warning is injected into the context via the Jinja template.
• Reduced performance degradation over long sessions — fewer server restarts needed.
• Optimized Jinja template — reduced nested conditionals for faster prompt processing.

📊 Inline Visualizer

• One-click download of visualizations (HTML or PNG).
• Streaming visualization support.
• Various performance improvements.

🌐 Tools & Uploads

• Weather tool fully rewritten for improved reliability.
• Check URL & IP optimized to prevent server slowdowns.
• YouTube tool removed — a more efficient replacement is in development.
• New file upload support for .ppt, .pptx, and .dart.

🆕 Minerve Heavy Preview

• Minerve Heavy Preview now available on the same enhanced system as Minerve Preview, bringing the new tool discovery system and reasoning improvements to the Heavy model tier.

📌 Misc

Various minor improvements and optimizations have also been applied.

Version 1.10 introduces Minerve-Preview, the memory system, and a series of major improvements covering search, images, security, and the user experience.

🧪 Minerve-Preview

• Minerve-Preview is now available: this is the pre-production version of Minerve, allowing new features to be tested before deployment.
• Memory system now available on Minerve-Preview.
• New patent search tools via EPO, HAL, Crossref, PubMed, arXiv, and Wikipedia, accessible through Minerve-Preview.

🔄 Proxy System

• Deployment of a proxy system with detailed stats, enhanced security, and failover/failback to Minerve in case of overload or unavailability, and vice versa.
• Currently deployed exclusively on Minerve-Preview.

🌐 Translation and Website

• Full translation of the website, FAQ, and Minerve blog posts.
• From now on, every update will be published in bilingual format (French and English).

📝 Token Capacity

• Output token capacity extended from 16,384 to 32,768 tokens.
• Minerve can now read and write files twice as large in a single block.

🖼️ Images

• Image search and retrieval: reintroduction of an improved version of image search and web page image extraction.
• Addition of a boolean argument to the web search tool to enable or disable image retrieval.
• Overhauled Wikipedia tool added to Minerve-Preview with a caching and image retrieval system.

🔒 Security

• Addition of a definitive security measure that prevents Minerve from calling tools without first closing the reasoning block.
• Addition of a new control mechanism to limit repetitions and prevent Minerve from getting stuck in an infinite loop.

🎲 Randomness and Regeneration

• Implementation of a random seed allowing Minerve to generate a different response for each request.
• You can now regenerate a request to obtain a distinct result.

👤 Accounts and Groups

• Complete rewrite of the account creation and Minerve group update system (more responsive, more reliable).
• Fixed a bug preventing user account creation and updates during the trial period.

🤗 Hugging Face

• Creation of the Minerve AI organization on Hugging Face.
• Publication of the first 8 models as well as Minerve's custom Jinja template, all under the MIT license.

🎨 Display and Ergonomics

• Renaming of web search tool names for improved clarity and accessibility.
• Improved display of quotes under Minerve's response.
• Improved display of metrics with the addition of reasoning token count: ⬆︎ 11,862 · ⬇︎ 644 · Σ 12,506 · 🧠 138 · 💾 11,242 · 📐 12.5k/262.1k (5%) · ⏱ 10.7s · ⚡ 125.7 t/s

📎 File Uploads

• Maximum file upload size in chat increased from 10 MB to 35 MB.

📌 Misc

Various minor improvements and changes have also been made.

Version 1.09 is a major update focused on stability, performance, and user experience.

🧠 System Prompt and Reasoning

• Complete rewrite of the system prompt to refine the structure and rephrase as many "do not..." as possible: several research papers and internal tests show that negations are detrimental.
• New improvement to the reasoning system to limit loops and reduce wasted time going through them.
• Refined regulation of hyperparameters to diversify responses and limit tool call loops.

🤖 Sub-Agent

• The number of tool calls per invocation increases from 10 to 30.

🔒 Enhanced Security

• Added security in the Jinja template to prevent Minerve from making a tool call before closing its thinking (source of crashes).
• Added several security measures to resolve issues of the LLM stopping abruptly without a response.
• Security measure in the template to avoid repeating a tool that has already failed (injection of a prompt inviting to use an alternative solution).

💬 Quota and Subscriptions

Complete improvement of the quota system:

• Alert message as soon as the quota reaches 100%, in addition to the quota exceeded message.
• Internal tagging system so that quota limit messages don't pollute the context once the quota is recharged.
• Custom messages depending on subscription type (Free / Starter / Pro).
• Minerve knows your exact quota.

🐳 Enriched Terminal

Added new libraries to Minerve's terminal: API creation and usage:

• fastapi, uvicorn, aiohttp
PHP website deployment:
• apache2, libapache2-mod-php, php8.3
DICOM medical image processing:
• pydicom — reading, modifying, and writing DICOM data (X-rays, MRIs, CT scans, etc.)
Image editing and conversion:
• Pillow (PIL Fork) — resizing, filtering, conversion, and pixel-by-pixel manipulation

🌐 Web Search and Cache

• Web search tool rewritten from scratch for URL retrieval.
• Resolution of a native Open WebUI issue regarding cache in requests using web search.
• Very strong speed improvement for all queries containing web tool usage.
• Thanks to extensive work on caching, the "cache hit" rate increased from 49% to 94% on average.

🖼️ Image Analysis

• The ceiling for simultaneous image analysis is raised from 8 to over 40 (240 in theory).

🌐 Website and FAQ

• Complete review of the website and blog system (subscription updates, technical FAQ, etc.).
• Slight design improvement.
• Major FAQ improvement: addition of many new details and clarifications.

📌 Miscellaneous

Many minor improvements and modifications have also been made.

Version 1.08 marks a new milestone in speed, stability, and research capabilities for Minerve.

🌐 Local Web Analysis

When you ask technical questions about a website, Minerve will now automatically dump the entire website and analyze it locally. This minimizes "I don't have access to the source code" responses.

💾 Backup System

Implementation of a backup system to minimize the potential loss of data in your shell space. Your files are now protected.

📊 R Scripts in the Terminal

Minerve can now execute and compile R scripts in its terminal. This opens up new possibilities for statistical analysis and scientific research.

📄 LibreOffice in the Terminal

All LibreOffice libraries have been added to Minerve's terminal:

• Creation and editing of .doc, .docx, .xls, .xlsx, .pptx, etc. files.
• Minerve can now generate and edit office documents directly.

🔒 Nmap for Cybersecurity

The Nmap library has been added to Minerve's terminal, enabling in-depth network scans and cybersecurity analysis.

🧠 System Prompt Revision

Complete update and revision of Minerve's system prompt:

• New changes
• Minor corrections
• Improved structure
• Clarification of certain ambiguities

⚡ Proxy Optimization

Additional proxy optimization for better streaming responsiveness across interfaces.

🐳 Harbor Server for Docker

Setup of a Harbor server to manage Docker images, providing additional optimization, security, and reliability.

🛡️ Security Measures and Optimizations

Several security measures and optimizations have been implemented:

• Tool calls: output limited to the first 20,000 characters
• File reading: output limited to the first 20,000 characters
• Processes and commands: output limited to the last 10 lines
• Log files: maximum size reduced from 10 MB to 1 MB

"This prevents the model from trying to read a file larger than its context and crashing. It will then understand that the file is too large and use other methods to analyze it."

These smart limits enable Minerve to better handle large files and automatically choose the most appropriate analysis method.

Version 1.07 brings significant gains in simplicity, speed, memory, and capabilities.

🧠 Memory System in Final Stages

The memory system, in development for several months, is now approaching completion. This major feature will allow Minerve to remember information between conversations.

"Ask me for early access on Discord if you're interested."

⚡ Instant File Upload

Major speed improvement when sending files in chat: no more embedding or retrieval. Even a large file is now instant, provided your internet connection allows it.

📁 File Organization Filter

Added a filter so that your files/images sent in chats are directly placed at the root of your shell. Better performance and capabilities for Minerve.

🔧 Automatic Server Startup

Added automatic startup of the 2 inference servers in case of power outage or server reboot, in addition to the automatic restart in case of a crash.

📈 Context Management Improvement

Improvement of the context management filter:

• Speed gain
• Better information retention over the long term in the same conversation
• Better optimization and compression

🔧 More Tool Calls

The maximum number of tool calls per turn increases from 100 to 150. Minerve can now perform even more complex actions in a single response.

🔥 Minerve Heavy Multi-Step

Improved speed of Minerve Heavy in multi-step, enabling even faster complex tasks.

Version 1.06 marks a new milestone in performance, reliability and autonomy for Minerve.

⚡ Reduced Energy Consumption

Minerve now consumes 33% less energy, with a negligible speed reduction of about 10%.

💬 Increased Messages per Minute

The number of messages per minute for free users increases from 2 to 4 — a 100% increase in interaction capacity.

🌐 Reflection Security

Added a safety mechanism to force Minerve to start with a reflection period. In some rare cases, Minerve could skip this step.

📅 Internal Knowledge Cutoff Date

Added the internal knowledge cutoff date for Minerve: this helps the model to doubt itself and search for recent information on the internet.

🧠 Minerve Sub-Agent

Implementation of Minerve Sub-Agent: Minerve now has a tool to launch autonomous sub-agents. When used well, the gain is tremendous.

✅ Automatic Restart

Implementation of an automatic restart in case of a crash on Minerve and Minerve Heavy (approximately 15 and 30 seconds of restart time).

📊 New Minerve Classic Quant

Change of the quant for Minerve Classic to a more compressed version: speed gain of 10-12%, bringing us back to -33% consumption and +2% speed.

🔍 Reddit and YouTube Domain Filtering

Filtering of www.reddit.com and youtube.com domains from Minerve's web search. Minerve has a dedicated tool for YouTube, which avoids polluting these search results. A list of filtered domains will be set up to improve search quality.

📈 Ngram Technology

Implementation of ngram technology for Minerve (a form of multi-token prediction): Minerve can gain 25 to 75% speed on specific tasks (code writing, repetitive documents).

🚀 MTP (Multi-Token Prediction) Technology

Implementation of MTP technology for Minerve Heavy: Minerve Heavy is 25-50% faster depending on the query.

📊 Visualization Tool

Added a visualization tool for Minerve (dynamic charts, etc. in the chat): just tell it "use your visualization tool".

🔧 Next Steps

"I will refine the settings for Minerve and Minerve Heavy in the coming weeks to make the most of these performance gains."

Version 1.05 marks a significant leap in web research performance and Minerve's infrastructure.

🌐 Ultra-Powerful Web Research

Minerve is now faster and more performant on all types of web research — it's common to see it retrieve 100 to 200 sources during a complex search.

🔌 UPS Upgrade

Minerve now lasts 25 minutes in case of a power outage, ensuring reliable service continuity.

⚡ Acquisition of a Server Rack

Minerve is now fully racked in its own server rack, marking an important step in the professionalization of the infrastructure.

🗣️ Speech-to-Text System

Implementation of a speech-to-text system: you can speak to Minerve and it can answer you orally.

"However, I currently advise against using it. To make it reliable, relevant, and fast, I need to invest an additional €750."

📊 Increased Sites Retrieved per Search

The maximum number of sites retrieved per web search increases from 3 to 10. Minerve can now explore 3.33 times more sites per tool call.

🚀 Web Loading Server (Playwright)

Added a web loading server for the fetch_url tool with Playwright:

• Minerve now gets a cleaner return when retrieving web page content
• It can load sites that require JavaScript
• It can bypass some bot detectors

🔧 Title Generation Model Change

Switch from Qwen3 4B to Gemma4 2B for title generation.

"Get ready for a surprise..."

Minerve continues to gain power with version 1.04. This update brings major improvements to performance, vision, and context management.

🐛 Critical Bug Fix

Fix of a bug introduced by the last update that prevented users from using the models ("model not found" error).

🏷️ Brand Identity

Added the text "Minerve" to web page titles and the homepage for better brand identification.

⚡ 100 Tool Calls per Response

Following significant gains in long multi-step tasks, Minerve goes from 30 maximum tool calls per response to 100. This increase allows for processing much more complex tasks in a single response.

🔥 Parallel Tool Calls

Minerve can now call tools in parallel, significantly accelerating the execution of tasks requiring multiple simultaneous tools.

🌐 Web Search and Scraping Speed

Improved web search and scraping speed for faster and more relevant results.

🖥️ Terminal Enabled by Default

Activation and connection to Minerve's terminal for everyone by default. Better capabilities, 2 fewer clicks for you.

🗜️ Vision Compression

Compression of the vision component of Minerve and Minerve Heavy from F16 to Q8, offering space gains with negligible loss of capability.

👁️ Vision Improvement

New improvement in vision capability with an increase in minimum tokens per image: 1024 → 2048. Minerve now analyzes images with twice as much detail.

🧠 Minerve Heavy Model Change

Switch to Qwen3.6-27B for Minerve Heavy, bringing a significant gain in code.

🧠 Minerve Model Change

Switch to Qwen3.6-35B for Minerve, also with a significant gain in code.

⚡ Context Management V6

Update of Minerve's context management system to V6, inspired by 2 research papers. This new version offers better context compression and overall gains in long conversations.

"This update solidifies Minerve's foundations and paves the way for future developments."

Version 1.03 brings significant performance improvements and new features to enhance the Minerve experience.

🔧 System Prompt Improvement

The system prompt has been optimized to reduce the number of tokens used by 40-50%. This optimization positively impacts Minerve's performance and responsiveness.

⚡ Cache Optimization

Prompt caching has been optimized for a notable performance gain, reducing response times and improving the user experience.

🔥 Minerve Model Optimization

The Minerve model has been significantly optimized: it goes from 100k to 262,144 tokens. This massive increase in the context window allows for processing much longer conversations and documents.

🧠 Model Change

The model behind Minerve has been updated to Qwen3.6-35B, bringing significant general gains, including a notable improvement in vision.

👁️ Vision Improvement

Minerve's vision capability has been improved, allowing for more accurate and detailed analysis of images and visual documents.

✅ Improved Quota System

The quota system has been improved with:

• A warning popup during intensive use
• Notifications when a quota is reached

🔧 Tool Call Optimization

The context of tool calls has been optimized, reducing the number of tokens by 40-50% per tool call.

🐛 Crash Fix

A potential crash affecting the Minerve server has been fixed, improving the overall stability of the platform.

📅 New Calendar System

A calendar and events system has been added for Minerve, with the ability to launch recurring tasks.

Example: "Give me a summary of the latest tech news every day at 8am"

This feature opens up new possibilities for automation and productivity for Minerve users.

Version 1.01 brings a series of major improvements and expands access for all users.

🖥️ Shell Access for Everyone

We have added full shell access for all users. Each account now has a Linux environment with:

• 6 CPU cores
• 8 GB of RAM
• 20 GB of storage

This opens up new possibilities for automation, development, and running scripts directly from the Minerve interface.

🔥 Minerve Heavy Open to All

The Minerve Heavy model is now available to all users. No more waiting — you can take advantage of its superior performance right now.

🔑 API Key Generation

API key generation is now open to everyone. Create your own keys to integrate Minerve into your own applications and workflows.

💰 New Pricing System

User tiers have evolved into a message-based and category-based pricing system:

"Minerve Heavy is 10 times more expensive than Minerve Classic, reflecting the superior computing power required."

This system ensures fair and transparent pricing, tailored to the actual use of each model.

🌐 Complete Website Rewrite in PHP

The website has been completely rewritten in PHP for 100% Google SEO optimization (excluding mobile performance). Minerve's natural search engine ranking is now optimal.

✅ Autonomous Skill Management

A new autonomous skill management tool has been added, inspired by Claude Code and Hermes Agent. Create, modify, and manage your own custom skills for Minerve.

📈 Minerve Classic Update

The Minerve Classic model has received major gains across all areas:

• Code: better generation and comprehension
• Prose: more fluid and natural writing
• Translation: increased accuracy
• Hallucination rate: significantly reduced
• Mathematics: more reliable calculations
• Logic: enhanced reasoning

We are proud to announce the official end of the Minerve beta test. Version 1.0 is now available, marking the transition from a testing phase to a fully operational platform.

🌐 Complete Website Redesign

The website has been completely redesigned and redeployed. A new architecture that is more modern, faster, and more intuitive to provide the best possible experience for our users.

💳 3-Tier Subscription System

We have implemented a structured subscription system with three plans:

• Free: Access essential features for free
• Starter: For regular users who need more power
• Pro: The complete package for professionals and teams

"Our goal is to make sovereign AI accessible to everyone while ensuring the sustainability of our infrastructure."

⚡ +12% GPU Bandwidth

The infrastructure has been optimized to deliver 12% additional GPU bandwidth, resulting in faster response times and increased fluidity for all interactions.

🔧 Coder Model → Heavy

The code-dedicated model has been replaced by Heavy, offering significantly better performance in code generation and comprehension.

🧠 Update of the 2 Main Models

Our two main models have been updated with major overall improvements:

• Better context understanding
• More precise and nuanced responses
• Enhanced logical reasoning

🖼️ Native Image Support

Both models now natively support images. Send us screenshots, diagrams, or visual documents and Minerve will analyze them directly.

✅ Welcome to Version 1.0

Minerve is no longer in beta. We are entering a new phase of development, with regular updates and increasingly ambitious features.

Thank you to the entire community for your participation in the beta test. Your feedback was essential to making this version 1.0 so powerful.

We are delighted to announce the official launch of our new blog platform! This development marks an important step in our communication strategy and information sharing with our community.

🎯 Why a Blog?

The "News" section on our homepage was perfect for quick updates, but we realized that our users needed more in-depth content. The blog allows us to:

• Share detailed tutorials: Learn how to use all of Minerve's features
• Document our developments: Follow the technical evolution of our AI
• Analyze complex topics: Dive into the heart of sovereign AI issues
• Create an accessible archive: Easily find all our past publications

📝 What to Expect

Our team is committed to regularly publishing quality content:

"Our mission is to provide a high-performance artificial intelligence while minimizing environmental impact, with full transparency and compliance with European regulations."

🔔 Stay Informed

Subscribe to our notifications so you don't miss any articles. We will cover:

• Technical updates and new features
• Analysis on AI and ecology
• Tutorials and user guides
• User feedback and case studies

We are proud to announce the launch of Minerve Coder, our new feature dedicated to developers and programming.

🛠️ Minerve Coder: An AI Specialized in Code

Minerve Coder has been specifically trained and optimized to understand and generate code in many programming languages. Whether you are a beginner or an expert developer, Minerve Coder is here to help you.

💳 Funding through Subscriptions

To maintain and improve our infrastructure, we have integrated Stripe as a payment solution. This allows us to:

• Finance the purchase of new GPUs to improve performance
• Maintain an eco-friendly and sustainable infrastructure
• Continuously develop new features
• Ensure the sovereignty of our platform

🚀 Next Steps

Thanks to your support, we plan to:

• Add new specialized language models
• Improve response times
• Develop collaborative features

A major investment to improve Minerve's performance: the addition of a third NVIDIA RTX 3090 GPU to our infrastructure.

📈 Impact on Performance

This new investment allows us to:

• Reduce response times by 40%
• Handle more complex queries
• Improve response quality
• Manage more simultaneous users

🔄 Reranking System

We have also implemented a new reranking system that allows us to:

"Prioritize the most relevant results and improve the accuracy of Minerve's responses"

💰 Funding

This investment was made with our own funds, demonstrating our commitment to the continuous improvement of the platform, even without subscription revenue.

An important step in the evolution of Minerve: the addition of a second 3090 GPU and a complete overhaul of PDF document processing.

📑 New PDF Engine

Our new PDF processing system offers:

• More accurate text extraction
• Layout preservation
• Table and image support
• Processing of documents hundreds of pages long

🌐 Optimized Website

The website has also benefited from significant optimizations:

• Load time reduced by 50%
• Better mobile experience
• More modern and accessible design

A major breakthrough for research: Minerve now integrates new scientific sources to enrich its responses and improve the quality of its analyses.

📚 New Sources

We have integrated several scientific and academic databases to offer more precise and better-documented answers.

🔍 Research Improvement

This new system enables Minerve to:

• Access recent scientific publications
• Cross-reference sources for enhanced validation
• Provide bibliographic references
• Improve the relevance of results

🎯 Objective

"Provide scientifically grounded answers, with verifiable sources and full transparency on the references used."

This update reinforces our commitment to the quality and reliability of the information provided by Minerve.