DeepSeek-R1-Distill-Llama-8B compressed using EVR-1 (Evrmind Reconstruction), a novel compression method developed independently by Evrmind. The compressed weights average approximately 3 bits per parameter; the total GGUF file (~3.9 GiB) includes additional metadata and structure overhead. A reasoning model that thinks step-by-step before answering.
EVR-1 Bafethu achieves 0.44% repetition at 500 tokens and 1.75% at 1000 tokens, maintaining coherent chain-of-thought reasoning while being nearly 4x smaller than F16.
~3.9 GiB | DeepSeek R1 Reasoning | Runs on laptops, desktops, and Android (Termux)
Note: HuggingFace may display an incorrect parameter count in the sidebar due to the custom compression format. EVR-1 is not a standard quantization (not Q2, Q3, Q4, etc).
Setup
You need two things: the model files (from this HuggingFace repo) and a platform binary (from GitHub).
Step 2: Download the binary for your platform from the Downloads table. Save the archive into the evr-1-bafethu-8b-reasoning directory, then extract it:
Network access (phone, tablet, other devices on the same WiFi):
./start-server.sh --network
The script will print the URL to open on other devices. The model runs on your computer; other devices just connect to the web UI. The --network and --cpu flags are only available in start-server.sh (Linux/macOS/Android).
See WEB_UI.md for more options and troubleshooting.
Quick Start (CLI)
These examples assume you have completed Setup and are in the repo directory.
The model weights (evr-deepseek-r1-llama-8b-reasoning.gguf, 3.93 GiB (~4.2 GB download)) are available from the Files tab on this HuggingFace page. Platform binaries are hosted on GitHub Releases. You can verify downloads with SHA256SUMS.txt.
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.
How Reasoning Works
The model uses DeepSeek R1's reasoning format. It first thinks through the problem internally, then provides a clean answer:
<think>
To find 15% of 240, I need to multiply 240 by 0.15.
240 x 0.15 = 36
</think>
15% of 240 is **36**.
Why EVR-1 Bafethu-8b-Reasoning?
This is a reasoning model compressed to under 4 GiB using EVR-1 (Evrmind Reconstruction). The model thinks through problems step-by-step using <think>...</think> tags before providing a final answer. Useful for math, logic, coding, and complex questions.
EVR-1 compresses the model to under 4 GiB while maintaining coherent chain-of-thought reasoning at 1000+ tokens. In our tests (5 continuation-style prompts), EVR-1 Bafethu achieved 0.44% repetition at 500 tokens and 1.75% at 1000 tokens.
Benchmarks
Coherence (lower is better)
Average 4-gram repetition rate, 5 continuation-style prompts:
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
EVR-1 Bafethu matches Q4_K_M perplexity while being 16% smaller (3.93 GiB vs 4.69 GiB). DeepSeek-R1-Distill-Llama-8B has higher perplexity on raw text benchmarks than the base Llama 3.1 8B, as expected for a model distilled for reasoning tasks.
Coherence tested with 5 continuation-style prompts at 500 and 1000 tokens each, temperature 0, no repeat penalty. See BENCHMARK_RESULTS.md for full coherence results and sample outputs.
Limitations
Context window has been tested up to 2048 tokens. Longer contexts may work but have not been validated at 3-bit compression.
Occasional minor character-level artefacts due to 3-bit compression.
Reasoning chains may occasionally be incomplete or circular.
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 (CUDA 12), Apple Silicon, or any Vulkan GPU
CPU-only: Supported but slower (use -ngl 0 or --cpu flag)
This model can generate incorrect, biased, or harmful content. Reasoning chains may contain errors or circular logic. Users should apply appropriate content filtering for user-facing applications. See MODEL_CARD.md for details.
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
This model is subject to three licenses:
Evrmind Free License 1.0: Covers the EVR-1 compression and distribution. Permits personal, research, and commercial use with attribution.
DeepSeek MIT License: Covers the DeepSeek R1 distillation. Permissive open-source license.
Llama 3.1 Community License: Covers the underlying Llama architecture. Permits commercial use for entities with fewer than 700 million monthly active users.
