theMIMER

MIMER Is Mythical Enhanced Reasoning

theMIMER is an intelligent, local-first AI assistant that runs entirely on your own hardware. It combines a local LLM with a Retrieval-Augmented Generation (RAG) knowledge base, multi-persona thinking strategies, vision, image generation, agentic tool use, and a built-in web chat interface.

Feed it your documentation, source code, PDFs, and web pages. Ask questions, analyze code, generate reports, find bugs, and get answers grounded in your actual data - not hallucinations.

Copyright 2026 Components4Developers - Kim Bo Madsen. See LICENSE for terms.

---

More information?

• Components4Developers - www.components4developers.com
• Components4Developers Blog - components4developers.blog

---

Key Features

• Local-first AI - Runs on your GPU with no cloud dependency. Your data never leaves your machine.
• RAG Knowledge Base - Ingest documents, source code, PDFs, and web pages. Answers are grounded in your actual content.
• Multi-model architecture - Use a small fast model for routing and a large capable model for answers. Optionally add external providers (Gemini, Claude, OpenAI, DeepSeek) for complex tasks.
• Multi-persona thinking - Critique, debate, code review, and meta-orchestration strategies where multiple LLM personas analyze your question from different angles.
• Vision - Describe, analyze, and compare images. Upload images in the web chat for multimodal conversations.
• Image generation - Generate images from text prompts using local Stable Diffusion.
• Agentic tools - The code agent can read, analyze, and modify source files. External tools (compiler, grep, test runner) extend its capabilities.
• Web chat UI - Built-in web server with a modern chat interface, file upload, image preview, streaming responses, and Mermaid diagram rendering.
• OpenAI-compatible API - Expose your local models as an OpenAI-compatible endpoint for integration with other tools.
• Self-improvement - Autonomous feedback loop that tests itself, finds weaknesses, and generates skills to improve.
• Fully configurable - Every aspect is controlled through well-documented INI configuration files.

---

Quick Start

1. Download theMIMER

Download the latest theMIMER.exe from the releases page.

2. Download required DLLs

theMIMER requires the llama.cpp shared libraries. Download the build matching your GPU:

llama.cpp (required):

• Go to: https://github.com/ggml-org/llama.cpp/releases
• Download the Windows release matching your GPU:
  • NVIDIA: llama-bXXXX-bin-win-cuda-cu12.x-x64.zip
  • AMD / Intel / Any GPU: llama-bXXXX-bin-win-vulkan-x64.zip
  • CPU only: llama-bXXXX-bin-win-avx2-x64.zip
• Extract and place the following DLLs in the same directory as theMIMER.exe:
  • llama.dll
  • ggml.dll
  • ggml-base.dll
  • ggml-cpu.dll
  • ggml-cuda.dll (NVIDIA CUDA build only)
  • ggml-vulkan.dll (Vulkan build only)
  • Any other ggml-*.dll files included in the release

GPU drivers:

• NVIDIA: Install the CUDA Toolkit matching your llama.cpp build (e.g., CUDA 12.x): https://developer.nvidia.com/cuda-toolkit Keep your GPU drivers up to date: https://www.nvidia.com/drivers
• AMD: Install the latest AMD Adrenalin drivers with Vulkan support: https://www.amd.com/en/support
• Intel: Install the latest Intel GPU drivers with Vulkan support: https://www.intel.com/content/www/us/en/download-center
• CPU only: No GPU drivers needed, but inference will be significantly slower

3. Download a chat model (GGUF)

You need at least one LLM model in GGUF format. Recommended starting models:

Model	Size	VRAM	Good for	Download
Qwen2.5-1.5B-Instruct (Q5_K_M)	~1.1 GB	~2 GB	Fast routing, reranking	HuggingFace
Qwen3.5-9B variant (Q8_0)	~9.5 GB	~12 GB	Chat, code, analysis, vision	See below

The recommended main model is a Qwen 3.5 9B variant with vision support. The variant used during development is Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q8_0.gguf with its matching CLIP projector mmproj-Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-BF16.gguf. Search HuggingFace for "Qwen3.5 9B GGUF" to find this or similar community variants that include both the model and mmproj file.

A single model handles chat, code, analysis, and vision, so you do not need separate models for different tasks.

Place the .gguf files in the same directory as theMIMER.exe (or use full paths in the config).

4. Download an embedding model (GGUF)

Required for the RAG knowledge base:

Model	Dimensions	Download
nomic-embed-text-v1.5 (Q8_0)	768	HuggingFace

Place in the same directory as theMIMER.exe.

5. Configure

Copy the sample configuration files to the theMIMER directory and rename them:

theMIMER.settings.ini
theMIMER.settings.providers.ini
theMIMER.settings.rules.ini
theMIMER.settings.tools.ini
theMIMER.settings.skills.ini
theMIMER.settings.strategies.ini
theMIMER.settings.selfimprove.ini

At minimum, edit theMIMER.settings.providers.ini:

• Set ModelPath in [LLMProvider.1] to your main model filename
• Set ModelPath in [LLMProvider.2] to your small routing model filename

And theMIMER.settings.ini:

• Set Model in [Embedding] to your embedding model filename
• Optionally configure [Sources] to point at your document and code folders

6. Run

theMIMER.exe

On first run, theMIMER will:

1. Load the embedding model
2. Load the LLM model(s)
3. Ingest any configured sources into the knowledge base
4. Start the web chat server

Open your browser to http://localhost:8080 and start chatting.

---

Optional: Vision

Vision enables theMIMER to understand and describe images. If you are using the recommended Qwen 3.5 9B variant, it already supports vision - you just need its matching CLIP projector file.

File	Purpose
Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q8_0.gguf	Vision-capable chat model (same as main model)
mmproj-Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-BF16.gguf	CLIP projector for the model

Both files are available from the same HuggingFace repository. Search for "Qwen3.5 9B Uncensored HauhauCS GGUF".

Additional DLL required for vision:

• mtmd.dll - Included in the llama.cpp release (multimodal support)

Configuration (in theMIMER.settings.providers.ini):

[LLMProvider.1]
ModelPath=Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Q8_0.gguf
ClipModelPath=mmproj-Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-BF16.gguf

And enable vision in theMIMER.settings.ini:

[Vision]
Enabled=1

---

Optional: Image Generation

theMIMER can generate images from text prompts using local Stable Diffusion.

Required:

• stable-diffusion.dll - Build from https://github.com/leejet/stable-diffusion.cpp (Download a pre-built release or compile with CUDA support)
• A Stable Diffusion model in .safetensors format, e.g.:
  • v1-5-pruned.safetensors (~4 GB VRAM) from HuggingFace

Place stable-diffusion.dll next to theMIMER.exe and configure in theMIMER.settings.ini:

[ImageGeneration]
Enabled=1
Provider=local
SDModel=v1-5-pruned.safetensors

---

Optional: External LLM Providers

theMIMER works fully offline with local models. Optionally, you can add external providers for tasks that benefit from larger models, such as complex analysis, evaluation, or self-improvement scoring.

Supported providers (configured in theMIMER.settings.providers.ini):

Provider	Type	Notes
Google Gemini	openai (compatible endpoint)	Very large context, competitive pricing
Anthropic Claude	anthropic	Excellent reasoning and evaluation
OpenAI GPT	openai	Strong all-rounder
DeepSeek	openai (compatible endpoint)	Strong at code, very affordable
Groq	openai (compatible endpoint)	Very fast inference

Set Enabled=1 and provide your API key for any provider you want to use. All external providers are disabled by default.

---

Directory Layout

After setup, your theMIMER directory should look like this:

theMIMER/
  theMIMER.exe                          - The executable
  LICENSE                               - License file
  llama.dll                             - llama.cpp core library
  ggml.dll                              - GGML backend library
  ggml-base.dll                         - GGML base operations
  ggml-cpu.dll                          - GGML CPU backend
  ggml-cuda.dll                         - GGML CUDA backend (NVIDIA GPU)
  ggml-vulkan.dll                       - GGML Vulkan backend (AMD/Intel/NVIDIA GPU)
  mtmd.dll                              - Multimodal support (optional, for vision)
  stable-diffusion.dll                  - Stable Diffusion (optional, for image gen)
  qwen2.5-7b-instruct-q8_0.gguf        - Main chat model (example)
  qwen2.5-1.5b-instruct-q5_k_m.gguf    - Small routing model (example)
  nomic-embed-text-v1.5-q8_0.gguf       - Embedding model
  mmproj-*.gguf                         - Vision CLIP projector (optional)
  v1-5-pruned.safetensors               - Stable Diffusion model (optional)
  theMIMER.settings.ini                 - Main configuration
  theMIMER.settings.providers.ini       - LLM provider registry
  theMIMER.settings.rules.ini           - Query rules and router extensions
  theMIMER.settings.tools.ini           - External tool definitions
  theMIMER.settings.skills.ini          - Thinking skill definitions
  theMIMER.settings.strategies.ini      - Multi-persona strategies
  theMIMER.settings.selfimprove.ini     - Self-improvement configuration
  knowledge.kbvs                        - Vector store (created on first ingest)
  uploads/                              - Per-session file workspace
  generated_images/                     - Generated images output
  skills/                               - Thinking skill .md files

---

Configuration Files

theMIMER is configured through seven INI files, each fully documented with inline comments explaining every setting, its valid values, and its defaults.

File	Purpose
theMIMER.settings.ini	Main settings: embedding, retrieval, orchestration, context, sampling, conversation, output, upload, chat server, vision, image generation, knowledge sources
theMIMER.settings.providers.ini	LLM provider registry: local models, external APIs, capabilities, selection strategies, query routing, cost controls
theMIMER.settings.rules.ini	Query rules: keyword/regex matching, intent forcing, scoped retrieval, router extensions
theMIMER.settings.tools.ini	External tools: compiler, grep, test runner, or any command-line tool the agent can invoke
theMIMER.settings.skills.ini	Thinking skills: code analysis, bug finding, refactoring, test generation, document writing, diagram generation
theMIMER.settings.strategies.ini	Multi-persona strategies: research, security, code analysis, translation teams with configurable personas
theMIMER.settings.selfimprove.ini	Self-improvement: autonomous testing, evaluation, gap detection, skill generation

---

Pipeline Architecture

theMIMER processes queries through a layered pipeline. Two settings control the depth:

Query Mode ([Query] Mode) sets the pipeline ceiling - which processing phases are available:

direct      Raw LLM only. No retrieval, no routing, no tools.
    |
retrieval   + RAG retrieval (vector + BM25 hybrid search)
    |
guided      + LLM router, skills, HyDE, query expansion
    |
agentic     + Tool use (code agent, upload agent) + query decomposition
    |
deep        + Multi-persona thinking strategies

Thinking Mode ([LLM] Thinking) controls how the LLM reasons - but only activates when Mode is deep:

Thinking	Description
off	Direct LLM response, no multi-persona reasoning
critique	Draft-then-critique: generate, review, refine
debate	Multiple personas argue perspectives, synthesizer merges
code	Specialized code analysis personas
auto	Auto-classify each query and pick the best strategy
meta	Iterative orchestrator: select, execute, score, refine

---

Runtime Commands

theMIMER supports slash commands in both the console and web chat. Type /help for a full list. Key commands:

Command	Description
/help	List all commands
/status	System status overview
/querymode <mode>	Set pipeline mode: direct, retrieval, guided, agentic, deep
/think <strategy>	Set thinking: off, critique, debate, code, auto, meta
/providers	Show provider status and statistics
/router show	Show router configuration
/skills	List active skills
/rules list	List active rules
/tools list	List registered external tools
/ingest	Ingest configured knowledge sources
/clear	Clear conversation history

---

System Requirements

• OS: Windows 10/11 (64-bit)
• GPU: Any GPU supported by llama.cpp (strongly recommended):
  • NVIDIA via CUDA (best performance, most tested)
  • AMD via Vulkan or ROCm
  • Intel via Vulkan or SYCL
  • 8 GB VRAM minimum for a 7B model
  • 12-16 GB VRAM recommended for larger models + vision
  • 24+ GB VRAM for 24B+ models
• CPU: Any modern x64 processor (CPU-only mode is supported but slow)
• RAM: 16 GB minimum, 32 GB recommended
• Disk: 10-20 GB for models and knowledge base

---

Links

• llama.cpp releases: https://github.com/ggml-org/llama.cpp/releases
• stable-diffusion.cpp: https://github.com/leejet/stable-diffusion.cpp
• NVIDIA CUDA Toolkit: https://developer.nvidia.com/cuda-toolkit
• NVIDIA Drivers: https://www.nvidia.com/drivers
• AMD Drivers: https://www.amd.com/en/support
• Intel Drivers: https://www.intel.com/content/www/us/en/download-center
• GGUF Models (HuggingFace): https://huggingface.co/models?sort=trending&search=gguf
• Nomic Embed (embedding model): https://huggingface.co/nomic-ai/nomic-embed-text-v1.5-GGUF
• Components4Developers: https://components4developers.blog

---

License

theMIMER is free to use for individuals and organizations. See LICENSE for the full terms, including the non-compete restriction.

Copyright 2026 Components4Developers - Kim Bo Madsen. All rights reserved.