Instructions to use Evrmind/EVR-1-Bafethu-8b-Reasoning with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Evrmind/EVR-1-Bafethu-8b-Reasoning with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Evrmind/EVR-1-Bafethu-8b-Reasoning
# Run inference directly in the terminal:
llama cli -hf Evrmind/EVR-1-Bafethu-8b-Reasoning

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Evrmind/EVR-1-Bafethu-8b-Reasoning
# Run inference directly in the terminal:
llama cli -hf Evrmind/EVR-1-Bafethu-8b-Reasoning

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Evrmind/EVR-1-Bafethu-8b-Reasoning
# Run inference directly in the terminal:
./llama-cli -hf Evrmind/EVR-1-Bafethu-8b-Reasoning

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Evrmind/EVR-1-Bafethu-8b-Reasoning
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Evrmind/EVR-1-Bafethu-8b-Reasoning

Use Docker

docker model run hf.co/Evrmind/EVR-1-Bafethu-8b-Reasoning

LM Studio
Jan

vLLM

How to use Evrmind/EVR-1-Bafethu-8b-Reasoning with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Evrmind/EVR-1-Bafethu-8b-Reasoning"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Evrmind/EVR-1-Bafethu-8b-Reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Evrmind/EVR-1-Bafethu-8b-Reasoning

Ollama
How to use Evrmind/EVR-1-Bafethu-8b-Reasoning with Ollama:
```
ollama run hf.co/Evrmind/EVR-1-Bafethu-8b-Reasoning
```
Unsloth Desktop
Docker Model Runner
How to use Evrmind/EVR-1-Bafethu-8b-Reasoning with Docker Model Runner:
```
docker model run hf.co/Evrmind/EVR-1-Bafethu-8b-Reasoning
```

Lemonade

How to use Evrmind/EVR-1-Bafethu-8b-Reasoning with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Evrmind/EVR-1-Bafethu-8b-Reasoning

Run and chat with the model

lemonade run user.EVR-1-Bafethu-8b-Reasoning-{{QUANT_TAG}}

List all available models

lemonade list

Atomic Chat

EVR-1-Bafethu-8b-Reasoning / MODEL_CARD.md

Evrmind

Upload folder using huggingface_hub

304bcee verified 6 months ago

preview code

Raw

History Blame Contribute Delete

7.69 kB

metadata

license: other
license_name: evrmind-free-1.0
license_link: LICENSE.md
language:
  - en
library_name: llama.cpp
tags:
  - deepseek
  - deepseek-r1
  - llama
  - llama-3.1
  - gguf
  - 3-bit
  - quantization
  - evr
  - evrmind
  - text-generation
  - reasoning
  - chain-of-thought
  - on-device
  - bafethu
pipeline_tag: text-generation
model-index:
  - name: Evrmind EVR-1 Bafethu-8b-Reasoning (DeepSeek R1 Distilled)
    results:
      - task:
          type: text-generation
        metrics:
          - name: Perplexity (wikitext-2, ctx=512)
            type: perplexity
            value: 14.4
          - name: Coherence (rep4 @ 500 tokens)
            type: repetition-rate
            value: 0.44
          - name: Coherence (rep4 @ 1000 tokens)
            type: repetition-rate
            value: 1.75

Evrmind EVR-1 Bafethu-8b-Reasoning (DeepSeek R1 Distilled), ~3.9 GiB

A custom 3-bit compression of DeepSeek-R1-Distill-Llama-8B that fits in under 4 GiB while maintaining coherent chain-of-thought reasoning at 1000+ tokens.

EVR-1 is not a standard quantization (not Q2, Q3, Q4, etc). It is a custom compression method developed by Evrmind. The compressed weights average approximately 3 bits per parameter; the total GGUF file (~3.9 GiB) includes additional metadata and structure overhead.

Overview

Chain-of-thought reasoning: The model thinks step-by-step using <think>...</think> tags before answering
Under 4 GiB: Runs on laptops, desktops, and edge devices (Android via Termux also supported)
Coherent at 1000+ tokens: Low repetition rate (1.75% rep4 at 1000 tokens)

How Reasoning Works

The model uses DeepSeek R1's reasoning format. It first thinks through the problem internally, then provides a clean answer:

User: What is the derivative of x^3 + 2x?

<think>
To find the derivative, I apply the power rule to each term:
- d/dx(x^3) = 3x^2
- d/dx(2x) = 2
So the derivative is 3x^2 + 2.
</think>

The derivative of x^3 + 2x is **3x^2 + 2**.

How to Run

Download the model file and the binary for your platform, then:

# Extract the binary
mkdir -p linux-cuda && tar xzf evrmind-linux-cuda.tar.gz -C linux-cuda

# Run (interactive chat with reasoning)
cd linux-cuda
LD_LIBRARY_PATH=. ./llama-cli -m ../evr-deepseek-r1-llama-8b-reasoning.gguf -ngl 99

# Run (single completion)
LD_LIBRARY_PATH=. ./llama-completion -m ../evr-deepseek-r1-llama-8b-reasoning.gguf -p "Your prompt here" -n 1000 -ngl 99

Platform Binaries

Platform	File	GPU Required
Linux + NVIDIA	`evrmind-linux-cuda.tar.gz`	NVIDIA GPU (CUDA 12)
Linux + Any GPU	`evrmind-linux-vulkan.tar.gz`	Any Vulkan-capable GPU
Windows + NVIDIA	`evrmind-windows-cuda.zip`	NVIDIA GPU (CUDA 12)
Windows + Any GPU	`evrmind-windows-vulkan.zip`	Any Vulkan-capable GPU
macOS (Apple Silicon)	`evrmind-macos-metal.tar.gz`	M1/M2/M3/M4
Android (Termux)	`evrmind-android-vulkan.tar.gz`	Vulkan

Note: The binaries are the same for the base, instruct, and reasoning models. You only need to download them once. Just point them at whichever GGUF you want to run.

Flags

Flag	Description
`-ngl 99`	Offload all layers to GPU (recommended)
`-n 1000`	Generate 1000 tokens (reasoning models need more tokens for thinking)
`-p "..."`	Your prompt
`-t 8`	Number of CPU threads (for CPU layers)

Model Details

Name: Evrmind EVR-1 Bafethu-8b-Reasoning (DeepSeek R1 Distilled)
Base model: DeepSeek-R1-Distill-Llama-8B (Llama 3.1 8B architecture)
Size: ~3.9 GiB (GGUF)
Method: EVR-1 (Evrmind Reconstruction), a custom 3-bit compression method
Backends: CUDA, Vulkan, Metal, CPU
Context: Tested up to 2048 tokens; longer contexts have not been validated at 3-bit compression
Chat template: DeepSeek R1 format (built-in)

Benchmarks

Coherence (5 continuation-style prompts, 500 and 1000 tokens each)

Average 4-gram repetition rate (lower = better):

Model	Size	rep4 @ 500	rep4 @ 1000
EVR-1 Bafethu	3.93 GiB	0.44%	1.75%

Perplexity

Model	Size	Perplexity (wikitext-2, ctx=512)
DeepSeek-R1-Distill-Llama-8B Q4_K_M	4.69 GiB	14.39
EVR-1 Bafethu	3.93 GiB	14.40

Also Available

EVR-1 Maano-8b, base model for text completion and creative writing
EVR-1 Maano-8b-Instruct, instruction-following chat model

Intended Use

This model is intended for on-device reasoning, math, logic, and coding tasks on laptops, desktops, and edge devices where memory is constrained. An Android (Termux) build is also available. There is no iOS build.

Limitations

Math reasoning quality is limited by the 3-bit compression level.
Occasional minor character-level artefacts (e.g., dropped letters) due to 3-bit compression.
Generation quality degrades somewhat beyond 1000 tokens.
Reasoning chains may occasionally be incomplete or circular.
The model may identify itself as "DeepSeek-R1"; this is expected, as the underlying model was trained by DeepSeek. A system prompt can be used to override this behaviour.
As with all heavily quantized models, generated text may contain factual inaccuracies (e.g., incorrect numbers, dates, or scientific details). Always verify factual claims independently.

System Requirements

Storage: ~4 GiB for model weights + ~50 MB for binaries
RAM: 6 GiB minimum (8 GiB recommended)
GPU (recommended): NVIDIA GPU with CUDA 12, Apple Silicon (M1/M2/M3/M4), or any Vulkan-capable GPU
CPU-only: Supported but significantly slower
OS: Linux (x86_64), macOS (Apple Silicon), Windows (x86_64), Android (Termux, ARM64)
Not supported: iOS, 32-bit systems

Safety and Responsible Use

This model inherits the capabilities and limitations of its base model (DeepSeek-R1-Distill-Llama-8B). Like all language models, it can generate incorrect, biased, or harmful content. Users should:

Not rely on this model for factual accuracy without verification
Not use this model to generate content that could cause harm
Apply appropriate content filtering for any user-facing applications
Be aware that 3-bit compression may amplify certain failure modes of the base model
Be aware that reasoning chains may contain errors or circular logic

Derivative Works

If you create derivative works, credit "EVR-1 Bafethu" in your model name and documentation. Commercial use is permitted subject to the Llama 3.1 Community License Agreement and DeepSeek MIT License.

License

Available for personal, research, and commercial use with attribution, subject to upstream license terms. See LICENSE.md for full terms.

Built with Llama. This model is a derivative of DeepSeek-R1-Distill-Llama-8B (MIT License) which is based on Meta's Llama 3.1 8B. Subject to the DeepSeek MIT License and the Llama 3.1 Community License Agreement in addition to the Evrmind license.

Citation

@misc{evrmind2026evr1bafethu8breasoning,
  title={Evrmind EVR-1 Bafethu-8b-Reasoning (DeepSeek R1 Distilled): A Custom 3-Bit Compression Method for Coherent On-Device Reasoning},
  author={Evrmind},
  year={2026},
  url={https://huggingface.co/evrmind/evr-1-bafethu-8b-reasoning}
}

Contact

Email: hello@evrmind.io
Issues: GitHub