Instructions to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with 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 winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M # Run inference directly in the terminal: llama cli -hf winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M # Run inference directly in the terminal: llama cli -hf winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
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 winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama cli -hf winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
Use Docker
docker model run hf.co/winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
- Llama new
- LM Studio
- Jan
- vLLM
How to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "winwinplatform/Kanana-2-1.3b-instruct-GGUF" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "winwinplatform/Kanana-2-1.3b-instruct-GGUF", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
- Ollama
How to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with Ollama:
ollama run hf.co/winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
- Unsloth Desktop
- Pi
How to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with Docker Model Runner:
docker model run hf.co/winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
- Lemonade
How to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
Run and chat with the model
lemonade run user.Kanana-2-1.3b-instruct-GGUF-Q4_K_M
List all available models
lemonade list
- Hermes Agent
How to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use winwinplatform/Kanana-2-1.3b-instruct-GGUF with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "winwinplatform/Kanana-2-1.3b-instruct-GGUF:Q4_K_M" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
Kanana-2-1.3B-Instruct GGUF (kanana2tiny)
Powered by Kanana
kakaocorp/kanana-2-1.3b-instruct๋ฅผ ์๋ณธ ๊ตฌ์กฐ(3:1 sliding-window/full attention ํ์ด๋ธ๋ฆฌ๋ + ๋ ์ด์ด๋ณ RoPE)๋ฅผ ๊ทธ๋๋ก ๋ณด์กดํด llama.cpp GGUF๋ก ๋ณํํ๊ณ imatrix ์์ํํ ๋ชจ๋ธ์ ๋๋ค. ์จ๋๋ฐ์ด์ค(Android/iOS) ์ถ๋ก ์ฉ์ ๋๋ค.
โ ๏ธ ํจ์น๋ llama.cpp๊ฐ ํ์ํฉ๋๋ค. ์ด GGUF์ ์ํคํ ์ฒ๋
kanana2tiny์ ๋๋ค. upstream llama.cpp์ ์ด๋ฅผ ๋ด์ฅํ ์ฑ(LM Studio, Ollama, PocketPal ๋ฑ)์๋ ์์ง ์ด ์ํคํ ์ฒ๊ฐ ์์ดunknown model architecture: 'kanana2tiny'์ค๋ฅ๊ฐ ๋ฉ๋๋ค. ํจ์น์ ๋น๋ ๋ฐฉ๋ฒ์ ์๋ ์ฌ์ฉ ๋ฐฉ๋ฒ์ ์ฐธ๊ณ ํ์ธ์.
ํ์ผ
๋ชจ๋ GGUF๋ ํจ์น๋ llama.cpp๋ก ๋ณํํ f16์์ ๊ฐ์ imatrix๋ก ์์ํํ์ต๋๋ค. ์๋ ์์น๋ ์ด ์ ์ฅ์์ ์ฌ๋ฆฐ ํ์ผ ๊ทธ๋๋ก ์ธก์ ํ ๊ฐ์ ๋๋ค.
| ํ์ผ | ํฌ๊ธฐ | ํ๊ตญ์ด 512 dPPL / top-1 | ํ๊ตญ์ด 4096 dPPL / top-1 | ์์ด 4096 dPPL / top-1 | ๋น๊ณ |
|---|---|---|---|---|---|
kanana2tiny-f16.gguf |
2.41 GiB | โ0 / 99.93% | โ0 / 99.96% | โ0 / 99.96% | ๊ธฐ์ค (์ฌ์์ํ์ฉ) |
kanana2tiny-Q8_0.gguf |
1,314 MB | +0.3% / 97.1% | +0.3% / 97.6% | +0.3% / 97.7% | |
kanana2tiny-Q6_K-im.gguf |
1,070 MB | +1.7% / 93.6% | +1.2% / 94.2% | +1.0% / 94.3% | ํ์ง ์ฐ์ |
kanana2tiny-Q6_K.gguf |
1,070 MB | +2.0% / 93.5% | +1.4% / 94.1% | +1.2% / 94.1% | |
kanana2tiny-Q5_K_M-im.gguf |
925 MB | +2.6% / 90.2% | +2.1% / 91.0% | +2.1% / 92.0% | ์ถ์ฒ |
kanana2tiny-Q5_K_M.gguf |
925 MB | +3.2% / 89.5% | +2.7% / 89.9% | +2.5% / 90.8% | |
kanana2tiny-Q4_K_M-im-emb8.gguf |
852 MB | +6.6% / 83.5% | +6.9% / 85.0% | +7.7% / 86.4% | ์๋ฒ ๋ฉ Q8_0 |
kanana2tiny-Q4_K_M-im.gguf |
814 MB | +6.9% / 83.1% | +7.2% / 84.8% | +8.1% / 86.1% | ์ฉ๋ ์ฐ์ |
kanana2tiny-Q4_K_M.gguf |
814 MB | +8.9% / 82.5% | +9.6% / 83.6% | +8.5% / 85.0% | |
kanana2tiny-IQ4_NL-im.gguf |
744 MB | +11.2% / 80.8% | +10.3% / 82.9% | +9.2% / 84.1% | |
kanana2tiny-Q4_0-im.gguf |
740 MB | +10.9% / 78.0% | +14.0% / 79.3% | +15.0% / 81.5% | ARM ํ๋กฌํํธ ์ฒ๋ฆฌ ์ต์ |
kanana2tiny-Q4_0.gguf |
739 MB | +14.9% / 76.5% | +18.8% / 77.2% | +18.4% / 79.6% | |
kanana2tiny-IQ4_XS-im.gguf |
718 MB | +11.4% / 80.8% | +11.0% / 82.8% | +9.6% / 83.9% | |
kanana2tiny-Q3_K_M-im.gguf |
682 MB | +24.8% / 70.7% | +31.3% / 72.4% | +32.0% / 74.5% | ๋น๊ถ์ฅ |
kanana2tiny.imatrix.gguf |
2 MB | โ | โ | โ | importance matrix |
-im= imatrix ์ ์ฉ. dPPL = HF ์๋ณธ(fp32) ๋๋น perplexity ์ฆ๊ฐ์จ, top-1 = ๋ค์ ํ ํฐ 1์์๊ฐ ์๋ณธ๊ณผ ๊ฐ์ ๋น์จ. KLD ๋ฑ ์ ์ฒด ์์น:results/quant/summary.txt- ์ธก์ : HF ์๋ณธ fp32 logits๋ฅผ
llama-perplexity --kl-divergence-baseํฌ๋งท์ผ๋ก ์ ์ฅํด ํ ํฐ ๋จ์ ๋น๊ต. ํ๊ตญ์ดยท์์ด ์ํค๋ฐฑ๊ณผ(20231101) ๊ฐ ์ฝ 3๋ง ํ ํฐ. - imatrix ๋ณด์ ๋ฐ์ดํฐ: ํ๊ฐ ๊ตฌ๊ฐ๊ณผ ๊ฒน์น์ง ์๋ ํ๊ตญ์ด/์์ด ์ํค๋ฐฑ๊ณผ + ์ฑํ ํ ํ๋ฆฟ QA ์ํ (161K ํ ํฐ).
- tied embedding(=์ถ๋ ฅ์ธต)์ ๊ธฐ๋ณธ์ ์ผ๋ก Q6_K๋ก ์ ์ง๋๋ค. ์ด ๋ชจ๋ธ์ 4bit์ ๋ฏผ๊ฐํด 5bitโ4bit์์ ์์ค์ด ๊ธ์ฆํ๋ฏ๋ก Q5_K_M-im์ ๊ถ์ฅํ๋ค.
๊ธฐ์กด ๊ณต๊ฐ GGUF(Qwen3 ๋งคํ)์์ ์ฐจ์ด
๊ธฐ์กด์ ๊ณต๊ฐ๋ ์ผ๋ถ GGUF๋ config๋ฅผ Qwen3ForCausalLM์ผ๋ก ๋ฐ๊ฟ ๋ณํํ ๊ฒ์
๋๋ค. ์ด ๋ฐฉ์์์๋ sliding window attention๊ณผ full-attention ์ธต์ YaRN(factor 40)์ด ์ฌ๋ผ์ ธ ์๋ณธ๊ณผ ๋ค๋ฅด๊ฒ ๋์ํฉ๋๋ค.
| ํ๊ตญ์ด PPL | ์๋ณธ | Qwen3 ๋งคํ Q4_K_M | ์ด ์ ์ฅ์ Q4_K_M-im | ์ด ์ ์ฅ์ Q5_K_M-im |
|---|---|---|---|---|
| ctx 512 | 16.57 | 19.64 (+19%) | +6.9% | +2.6% |
| ctx 4096 | 14.09 | 31.68 (+125%) | +7.2% | +2.1% |
| ctx 16384 | 13.25 | ๋ถ๊ดด | โ | โ |
YaRN์ ์งง์ ๋ฌธ๋งฅ์์๋ attention ์ค์ผ์ผ(ร1.369)๊ณผ ์ ์ฃผํ RoPE ์ฐจ์์ ๋ฐ๊ฟ๋๋ค. ๊ทธ๋์ ๋งคํ ๋ฐฉ์์ 1024ํ ํฐ ์ด๋ด์์๋ ์๋ณธ๊ณผ ๋ค๋ฆ ๋๋ค(top-1 ์ผ์น์จ ์ฝ 82%). ํจ์น๋ f16 GGUF๋ HF ์๋ณธ๊ณผ KLD โ 1e-6, top-1 ์ผ์น์จ 99.9% ์ด์์ ๋๋ค(ctx 512~16K).
์ฌ์ฉ ๋ฐฉ๋ฒ
1. ํจ์น๋ llama.cpp ๋น๋
๊ธฐ์ค ์ปค๋ฐ 4da6337767f973e2b4d0797e5b323d77d8565e4a์ kanana2tiny ์ํคํ
์ฒ ํจ์น๋ฅผ ์ ์ฉํฉ๋๋ค.
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
git checkout 4da6337767f973e2b4d0797e5b323d77d8565e4a
git am 0001-model-add-Kanana-2-Tiny-kanana2tiny-architecture.patch
cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
2. ์คํ
./build/bin/llama-cli -m kanana2tiny-Q5_K_M-im.gguf --jinja -c 4096 -t 4 -fa off -lm none
--jinja: Kanana ์ฑํ ํ ํ๋ฆฟ์ ์ ์ฉํฉ๋๋ค. ๊ธฐ๋ณธ๊ฐ์thinking_mode = no_think์ ๋๋ค.- ๋ชจ๋ฐ์ผ CPU ๊ถ์ฅ ์ต์
(Galaxy S20 ์ค์ธก ๊ธฐ์ค):
-t <์ฑ๋ฅ ์ฝ์ด ์>-fa off: ๋๋ฉด ๊ธด ๋ฌธ๋งฅ ์์ฑ ์๋๊ฐ +58%์ ๋๋ค.-lm none: mmap์ ๋๋๋ค. ๊ธฐ๋ณธ mmap์ CPU repack์ด ๊ฒน์น๋ฉด ๊ฐ์ค์น๊ฐ ๋ฉ๋ชจ๋ฆฌ์ ์ด์ค์ผ๋ก ์ฌ๋ผ๊ฐ์, ๋๋ฉด ๋ฉ๋ชจ๋ฆฌ๊ฐ -31%์ ๋๋ค.
- ์ฑ์ ์๋ฒ ๋ํ ๋ llama.cpp C API์ ๊ธฐ๋ณธ๊ฐ
swa_full=true๋false๋ก ๋ฐ๊พธ๊ธฐ๋ฅผ ๊ถ์ฅํฉ๋๋ค. KV ์บ์๊ฐ ์ฝ 240MB ์ค์ด๋ญ๋๋ค. - ๋ฉํฐํด์ ์ง์ ๊ตฌํํ๋ค๋ฉด ํ
ํ๋ฆฟ ์ธ์
preserve_thinking=true๋ก ํฌ๋งทํ์ธ์. ์ด ํ ํ๋ฆฟ์ ์ด์ assistant ํด์์<think></think>๋ฅผ ๋นผ๊ณ ๋ ๋๋งํ๊ธฐ ๋๋ฌธ์, ๊ณผ๊ฑฐ ํฌ๋งท๊ณผ์ ์ฐจ์ด๋ง ์๋ผ ๋ถ์ด๋ ๋ฐฉ์์ด ์ด๊ธ๋ฉ๋๋ค.
์ค๊ธฐ๊ธฐ ์ฐธ๊ณ ์น (CPU, Q5_K_M-im, ctx 4096)
| ๊ธฐ๊ธฐ | ์งง์ ๋ํ ์์ฑ | 1.5Kํ ํฐ ์ ๋ ฅ ํ ์์ฑ | ๋ฉ๋ชจ๋ฆฌ |
|---|---|---|---|
| Galaxy S20 (Snapdragon 865, 4์ค๋ ๋) | 18.6 tok/s | 11.9 tok/s | ์ฝ 1.55 GB |
| Galaxy S23 Ultra (Snapdragon 8 Gen 2, 4์ค๋ ๋, ์ํ ์ฑ) | ์ฝ 18~19 tok/s | โ (๋ฏธ์ธก์ ) | ์ฝ 1.52 GB |
์์ ๋ด์ญ (Modifications)
- ์๋ณธ BF16 ๊ฐ์ค์น๋ฅผ ํจ์น๋ llama.cpp ๋ณํ๊ธฐ(
Kanana2TinyForCausalLMโkanana2tiny)๋ก f16 GGUF๋ก ๋ณํํ์ต๋๋ค. ๊ฐ์ค์น ๊ฐ ์์ฒด๋ ์์ ํ์ง ์์์ต๋๋ค. - imatrix๋ฅผ ์์ฑํด Q4_K_M / Q5_K_M / Q6_K๋ก ์์ํํ์ต๋๋ค. tied embedding(=์ถ๋ ฅ์ธต)์ Q6_K๋ฅผ ์ ์งํ์ต๋๋ค.
- GGUF ๋ฉํ๋ฐ์ดํฐ
general.name์ "Kanana 2 1.3b Instruct"๋ก ์ค์ ํ์ต๋๋ค. - ๋ณํยท์์ํ ํ๊ฒฝ: x86_64 Linux (llama.cpp CPU ๋น๋). ARM(Mac)์์ ๊ฐ์ ์ ์ฐจ๋ก ์์ํํ๋ฉด ๋นํธ ๋จ์๋ก๋ ๋ฌ๋ผ์ง์ง๋ง ํ์ง์ ์ค์ฐจ ๋ฒ์ ๋ด ๋์ผํจ์ ํ์ธ(Q5_K_M-im: KLD 0.0275 vs 0.0269)
- ์์ ์: winwinplatform, 2026-09
๋ผ์ด์ ์ค ๋ฐ ์ฌ์ฉ ์ ํ
์ด ๋ชจ๋ธ์ Kanana Open License Agreement์ ํ์๋ฌผ(Derivative Works)์ ๋๋ค. ์ ๋ฌธ์ ๋๋ด๋ LICENSE์ NOTICE๋ฅผ ์ฐธ๊ณ ํ์ธ์.
- Kanana is licensed in accordance with the Kanana Open License Agreement. Copyright ยฉ KAKAO Corp. All Rights Reserved.
- ์ฌ์ฉ ์ ๊ด๋ จ ๋ฒ๋ น๊ณผ KAKAO์ Guidelines For Responsible AI(๊ธ์ง ์ฌ์ฉ ์ ์ฑ , License ยง2.2)๋ฅผ ๋ฐ๋ผ์ผ ํฉ๋๋ค. ์ฌ๋ฐฐํฌํ ๋๋ ํ์ ์ฌ์ฉ์์๊ฒ๋ ๊ฐ์ ์ ํ์ด ์ ์ฉ๋จ์ ์๋ ค์ผ ํฉ๋๋ค.
- ๋ค์ ๊ฒฝ์ฐ์๋ KAKAO์ ๋ณ๋ ์์
๋ผ์ด์ ์ค๊ฐ ํ์ํฉ๋๋ค(License ยง4.1).
- APIยทํด๋ผ์ฐ๋๋ก ์ 3์์๊ฒ ์ ๊ณตํ๋ ๊ฒฝ์ฐ
- SIยท์จํ๋ ๋ฏธ์ค๋ก ๋ฉํํ๋ ๊ฒฝ์ฐ
- ์จ๋๋ฐ์ด์ค ์ ํ์ ์๋ฒ ๋ํด ์ 3์์๊ฒ ์ ๊ณตยทํ๋งคํ๋ ๊ฒฝ์ฐ
๊ฒฐ๊ณผ ํ์ผ (results/)
| ๊ฒฝ๋ก | ๋ด์ฉ |
|---|---|
results/quant/summary.txt |
์์ํ 13์ข ์์ฝํ (dPPL, KLD, top-1, 99% KLD) |
results/quant/*.log |
๋ณํ๋ณ llama-perplexity KLD ์์ ๋ก๊ทธ (ko 512 / ko 4096 / en 4096) |
results/eval/hf_variants.json |
HF์์ YaRN/SWA ์ ๊ฑฐ ๋ณํ์ PPLยทKLD (Qwen3 ๋งคํ ์์ค ๋ถ์) |
results/eval/kld_k2f16_*.log |
ํจ์น f16 GGUF vs HF ์๋ณธ ๊ฒ์ฆ (ctx 4096, 16384) |
results/eval/ppl_*.log, kld_qwen3map-*.log |
Qwen3 ๋งคํ GGUF(๊ธฐ์กด ๊ณต๊ฐ ๋ฐฉ์)์ ๊ณต๊ฐ Q4_K_M ์ธก์ |
results/device/ |
Galaxy S20 ์ค์ธก (์ค๋ ๋, 4์ข ์๋, ๊ธด ๋ฌธ๋งฅ, FA/KV, mmapยทrepack๋ณ ๋ฉ๋ชจ๋ฆฌ) |
์ฒดํฌ์ฌ
SHA256SUMS ์ฐธ๊ณ .
- Downloads last month
- 235
Model tree for winwinplatform/Kanana-2-1.3b-instruct-GGUF
Base model
kakaocorp/kanana-2-1.3b-base