How to use from
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 KorshLAB/KorshAI-2.5-6B:Q6_K
# Run inference directly in the terminal:
llama cli -hf KorshLAB/KorshAI-2.5-6B:Q6_K
Install from WinGet (Windows)
winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf KorshLAB/KorshAI-2.5-6B:Q6_K
# Run inference directly in the terminal:
llama cli -hf KorshLAB/KorshAI-2.5-6B:Q6_K
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf KorshLAB/KorshAI-2.5-6B:Q6_K
# Run inference directly in the terminal:
./build/bin/llama cli -hf KorshLAB/KorshAI-2.5-6B:Q6_K
Use Docker
docker model run hf.co/KorshLAB/KorshAI-2.5-6B:Q6_K
Quick Links

KorshAI 2.5 6B Banner

KorshAI 2.5 (6B GGUF)

KorshAI 2.5 6B is a proprietary, frontier-grade 6-billion parameter compact foundation coding, systems engineering, and modern web UI model architected and trained by KorshLab and the Korsh Foundation.

Engineered specifically for low-level systems programming, surgical code refactoring via SEARCH/REPLACE diff blocks, deep architectural Chain-of-Thought (<thought> ... </thought>), and cutting-edge 2026 frontend design (Tailwind CSS, OKLCH color palettes, Apple Glass glassmorphism, and Bento Grids).


๐ŸŒŸ Key Capabilities & Pillars

1. Structured Chain-of-Thought Reasoning (<thought>)

Before emitting code or solutions, KorshAI systematically explores edge cases, invariant validation, memory bounds (RAII / borrow checker), and concurrency semantics inside structured <thought> blocks.

2. Surgical Code Refactoring (SEARCH / REPLACE)

KorshAI produces production-grade diffs without lazy AI tropes (// TODO, empty placeholders, or rewriting whole 1,000-line files): ```markdown <<<<<<< SEARCH func (s *OrderService) ExecuteTrade(ctx context.Context, trade *TradeRequest) error { account, err := s.repo.GetAccount(ctx, trade.AccountID)

func (s *OrderService) ExecuteTrade(ctx context.Context, trade *TradeRequest) error { account, err := s.repo.GetAccountWithLock(ctx, trade.AccountID)

REPLACE


### 3. Polyglot Systems Engineering
Compiler-vetted across:
- **C11 / C++20:** Memory safety, RAII, zero undefined behavior, lock-free queues, cache alignment.
- **Rust:** Strict borrow checker compliance, lifetime management, Tokio async concurrency.
- **Go:** Goroutine leak prevention, mutex & atomic CAS patterns, context cancellation.
- **Python:** High-throughput async pipelines, FastAPI, Pydantic strict schemas.
- **TypeScript:** Strict type checking, zero `any`, algebraic data types.

### 4. Modern 2026 UI/UX Craft
Specialized in clean, non-generic frontend layouts:
- Deep obsidian dark mode (`#09090B`).
- Micro-borders (`border-white/10`).
- Backdrop filters (`backdrop-blur-xl`).
- Dynamic metrics in `JetBrains Mono` and typography in `Inter`.
- Realistic data schemas (zero *Lorem Ipsum*).

---

## ๐Ÿ“Š Benchmark & Inference Velocity

| Runtime / Quant | Prompt Processing (tok/s) | Generation Velocity (tok/s) | Recommended VRAM |
| :--- | :---: | :---: | :---: |
| **Q6_K (3.32 GB)** | **1,800.6 tok/s** | **88.3 tok/s** | **6 GB - 8 GB** |
| **Q8_0 (4.29 GB)** | **1,720.0 tok/s** | **82.5 tok/s** | **8 GB** |

---

## ๐Ÿš€ Quickstart Guide

### 1. Run with `llama.cpp` CLI
```bash
llama-cli -m KorshAI-2.5-6B-Q6_K.gguf \
  -p "<|im_start|>system\nEres KorshAI 2.5 6B, el modelo fundacional de inteligencia artificial de KorshLab.<|im_end|>\n<|im_start|>user\nDiseรฑa un mutex lock-free en C11 para una cola concurrente.<|im_end|>\n<|im_start|>assistant\n<thought>\n" \
  -n 2048 -c 8192 --temp 0.2 -ngl 99

2. Run with llama-server (OpenAI Compatible API)

llama-server -m KorshAI-2.5-6B-Q6_K.gguf \
  --host 0.0.0.0 --port 8080 -ngl 99 -c 8192 --metrics

3. Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="KorshAI-2.5-6B-Q6_K.gguf",
    n_gpu_layers=-1,
    n_ctx=8192,
)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "Eres KorshAI 2.5 6B, el modelo fundacional de KorshLab."},
        {"role": "user", "content": "Explica cรณmo evitar condiciones de carrera en Go con bloqueos pesimistas."}
    ],
    temperature=0.2,
)
print(response["choices"][0]["message"]["content"])

๐Ÿ›ก๏ธ License

Released under the Apache 2.0 License by KorshLab and Korsh Foundation.

Downloads last month
-
GGUF
Model size
4B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

6-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support