---
license: apache-2.0
language:
  - ru
  - en
tags:
  - text-generation
  - gguf
  - llama.cpp
  - llama-cpp-python
  - ollama
  - moe
  - mixture-of-experts
  - conversational
  - chat
  - assistant
  - instruction-following
  - large-language-model
  - llm
  - quantized
  - mxfp4
  - q8_0
  - 4bit
  - multimodal-text
  - multilingual
  - russian
  - english
  - local
  - offline
  - free
  - inference
  - deployment
  - transformers
pipeline_tag: text-generation
library_name: llama.cpp
---

# Blind Text Models

A collection of language models from the **Blind** family. Named for its focus on text tasks — from conversation to content generation.

## Current Model (single entry)

| Model | Parameters | Quantization | File size | Context | Focus |
|---|---|---|---|---|---|
| **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 |

Only one model is in the collection for now. New versions will be added to this table.

## Architecture

- Base architecture: **in-house** (Blind series)
- 24 blocks, embedding size 2880
- 32 experts, 4 active per token
- Context up to **128K tokens**
- Quantization: MXFP4 (expert weights) + Q8_0 (embeddings/output)
- Format: GGUF, runs via llama.cpp / Ollama / llama-cpp-python

## What It Is Good At

- **Lively conversation** — natural answers, casual and business communication in Russian and English
- **Explanations** — breaks down complex topics clearly
- **Text generation** — letters, posts, articles, concise summaries
- **Instructions and questions** — holds context well, knows how to clarify and answer to the point

> The model introduces itself as **Blind 1** — that is how it presents itself when asked. This is a build feature.

## Hardware / VRAM

Runs 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.

## Benchmarks

Instrumental metrics (MMLU and similar) are still being measured and will be added here. Generation speed is already benchmarked:

| Environment | Generation |
|---|---|
| NVIDIA RTX 5080 (16GB), llama.cpp | ≈ **10 tok/s** (100 tokens in ~10 s, including the reasoning prefix) |

Parameter estimates:
- Total parameters: **20 914 757 184 (~20.9B)** — counted from the model tensors
- Parameters excluding embeddings and output layer: **19.76B**

## Usage

```bash
# llama.cpp
llama-cli -m blind-1.gguf -p "Hello, who are you?"
```

```python
# llama-cpp-python
from llama_cpp import Llama
llm = Llama(model_path="blind-1.gguf", n_ctx=8192, n_gpu_layers=-1)
print(llm.create_chat_completion(messages=[{"role": "user", "content": "Who are you?"}]))
```

## License

The model is distributed under the Apache 2.0 license.