Llama 3.1 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 (3.93 GiB) includes additional metadata and structure overhead.
At 3.93 GiB, standard quantizations like Q3_K_M collapse into repetition. EVR-1 produces coherent text at 1000+ tokens where others fail: 5.83% repetition vs 77% for Q3_K_M at the same size.
3.93 GiB | Llama 3.1 8B base | 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 1: Clone this repo or download the files:
bash
1# Option A: Clone everything (~4.2 GB, requires git-lfs)2git lfs install3git clone https://huggingface.co/evrmind/evr-1-maano-8b
4cd evr-1-maano-8b
56# Option B: Or download individual files from the "Files" tab above
Step 2: Download the binary for your platform from the Downloads table. Save the archive into the evr-1-maano-8b 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. Note: 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 evr-1-maano-8b directory.
Linux + NVIDIA GPU:
bash
1cd linux-cuda
2LD_LIBRARY_PATH=. ./llama-completion -m ../evr-llama-3.1-8b.gguf -p "The main causes of the French Revolution were" -n 500 -ngl 99
The model weights (evr-llama-3.1-8b.gguf, 4.22 GB) are available from the Files tab on this HuggingFace page. Platform binaries are hosted on GitHub Releases. You can verify downloads with SHA256SUMS.txt.
Why EVR-1 Maano-8b?
Standard quantizations at 3-4 GiB collapse into repetition during extended generation. Here is one example from our test set (see BENCHMARK_RESULTS.md for all results):
Q3_K_M (3.83 GiB):
"The process of nuclear fusion in stars begins when the core of a star is hot enough to start fusing hydrogen into helium... The process of nuclear fusion in stars is a complex process... The process of nuclear fusion in stars is a complex process..." (loops indefinitely)
EVR-1 Maano-8b (3.93 GiB):
"The process of nuclear fusion in stars begins when the core of the star is made up of hydrogen... In stars like our sun, hydrogen atoms fuse together to form helium atoms. The helium atoms then fuse together forming carbon-12..." (continues coherently for 500+ words)
Benchmarks
Head-to-head, same base model (Llama 3.1 8B), different compression methods. All models tested with the same binary, temperature 0, no repeat penalty.
Coherence (lower is better)
Generation length
EVR-1 Maano (3.93 GiB)
Q3_K_M (3.83 GiB)
Q4_K_M (4.69 GiB)
5 prompts @ 500 tokens
5.83% rep4
76.79%
79.45%
5 prompts @ 1000 tokens
19.68% rep4
87.65%
89.69%
EVR-1 maintains coherent output where Q3_K_M and Q4_K_M degrade into repetition. This is the core advantage of EVR-1 at this compression level.
Accuracy
Benchmark
EVR-1 Maano (3.93 GiB)
Q3_K_M (3.83 GiB)
Q4_K_M (4.69 GiB)
ARC-Challenge (25-shot, 1172q)
59.8%
60.8%
61.3%
Perplexity (wikitext-2, ctx=512)
6.70
7.02
6.58
Perplexity (wikitext-2, ctx=2048)
6.19
6.13
5.74
Perplexity varies with context size. At the default context (512), EVR-1 outperforms Q3_K_M. At extended context (2048), all three are closer. See BENCHMARK_RESULTS.md for full methodology, raw outputs, and additional evaluations.
Limitations
This is a base model (not instruction-tuned). It completes text, but does not follow instructions or engage in conversation without additional prompting. The web UI works for open-ended generation and experimentation.
Context window has been tested up to 2048 tokens. Longer contexts may work but have not been validated at 3-bit compression.
Perplexity at default context (512) is 6.70, outperforming Q3_K_M (7.02) but higher than Q4_K_M (6.58). At extended context (2048): 6.19 vs Q3_K_M 6.13, Q4_K_M 5.74.
Occasional minor character-level artefacts due to 3-bit compression.
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. It has not been safety-tuned or RLHF-aligned. 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 Maano" in your model name and documentation. Commercial use is permitted subject to the Llama 3.1 Community License Agreement.
License
This model is dual-licensed:
Evrmind Free License 1.0: Covers the EVR-1 compression and distribution. Permits personal, research, and commercial use with attribution.
Llama 3.1 Community License: Covers the underlying Llama 3.1 weights. Permits commercial use for entities with fewer than 700 million monthly active users.
