debugdll/Blind-Text-Models

1---2license: apache-2.03language:4  - ru5  - en6tags:7  - text-generation8  - gguf9  - llama.cpp10  - llama-cpp-python11  - ollama12  - moe13  - mixture-of-experts14  - conversational15  - chat16  - assistant17  - instruction-following18  - large-language-model19  - llm20  - quantized21  - mxfp422  - q8_023  - 4bit24  - multimodal-text25  - multilingual26  - russian27  - english28  - local29  - offline30  - free31  - inference32  - deployment33  - transformers34pipeline_tag: text-generation35library_name: llama.cpp36---37 38# Blind Text Models39 40A collection of language models from the **Blind** family. Named for its focus on text tasks — from conversation to content generation.41 42## Current Model (single entry)43 44| Model | Parameters | Quantization | File size | Context | Focus |45|---|---|---|---|---|---|46| **Blind Text 1** (`blind-1.gguf`) | **20.9B** (MoE, 32 experts, 4 active) | MXFP4 + Q8_0 | 11.3 GB | 131 072 (128K) | All-purpose assistant, strong at conversation, explanations and generation in Russian and English |47 48Only one model is in the collection for now. New versions will be added to this table.49 50## Architecture51 52- Base architecture: **in-house** (Blind series)53- 24 blocks, embedding size 288054- 32 experts, 4 active per token55- Context up to **128K tokens**56- Quantization: MXFP4 (expert weights) + Q8_0 (embeddings/output)57- Format: GGUF, runs via llama.cpp / Ollama / llama-cpp-python58 59## What It Is Good At60 61- **Lively conversation** — natural answers, casual and business communication in Russian and English62- **Explanations** — breaks down complex topics clearly63- **Text generation** — letters, posts, articles, concise summaries64- **Instructions and questions** — holds context well, knows how to clarify and answer to the point65 66> The model introduces itself as **Blind 1** — that is how it presents itself when asked. This is a build feature.67 68## Hardware / VRAM69 70Runs fully on GPU in ~11.5 GB — fits comfortably in a 12 GB VRAM card, and easily on 16 GB+. CPU-only inference works too (slower). No external API keys or cloud required — fully local and private.71 72## Benchmarks73 74Instrumental metrics (MMLU and similar) are still being measured and will be added here. Generation speed is already benchmarked:75 76| Environment | Generation |77|---|---|78| NVIDIA RTX 5080 (16GB), llama.cpp | ≈ **10 tok/s** (100 tokens in ~10 s, including the reasoning prefix) |79 80Parameter estimates:81- Total parameters: **20 914 757 184 (~20.9B)** — counted from the model tensors82- Parameters excluding embeddings and output layer: **19.76B**83 84## Usage85 86```bash87# llama.cpp88llama-cli -m blind-1.gguf -p "Hello, who are you?"89```90 91```python92# llama-cpp-python93from llama_cpp import Llama94llm = Llama(model_path="blind-1.gguf", n_ctx=8192, n_gpu_layers=-1)95print(llm.create_chat_completion(messages=[{"role": "user", "content": "Who are you?"}]))96```97 98## License99 100The model is distributed under the Apache 2.0 license.