Kanana 2 3B Instruct โ€” GGUF (Q4_K_M)

Powered by Kanana

kakaocorp/kanana-2-3b-instruct ๋ฅผ llama.cpp ์—์„œ ๋ฐ”๋กœ ์“ธ ์ˆ˜ ์žˆ๊ฒŒ GGUF ๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  4bit(Q4_K_M) ์–‘์žํ™”ํ•œ ํŒŒ์ผ์ž…๋‹ˆ๋‹ค. ์›๋ณธ ๋ฐฐํฌ์—๋Š” GGUF ๊ฐ€ ์—†์–ด ์‚ฌ๋‚ด ์˜จ๋””๋ฐ”์ด์Šค(๋กœ์ปฌ LLM) ์šฉ๋„๋กœ ์ง์ ‘ ๋ณ€ํ™˜ํ–ˆ์Šต๋‹ˆ๋‹ค.

ํŒŒ์ผ kanana-2-3b-instruct-Q4_K_M.gguf
ํฌ๊ธฐ 2,161,793,408 bytes (2.01 GiB)
์–‘์žํ™” Q4_K_M โ€” 4.92 BPW (์›๋ณธ BF16 16.00 BPW)
์•„ํ‚คํ…์ฒ˜ qwen3 (์›๋ณธ config.json ์ด Qwen3ForCausalLM)
ํ† ํฌ๋‚˜์ด์ € tokenizer.ggml.pre = kanana2, vocab 128,256
์ปจํ…์ŠคํŠธ 32,768 (rope: yarn, factor 40, original 4,096)
๋Œ€ํ™” ์„œ์‹ ์›๋ณธ chat_template.jinja ๋ฅผ GGUF ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ์— ํฌํ•จ

๋ณ€๊ฒฝ ์‚ฌํ•ญ (Kanana Open License ยง3.1(iii))

๊ฐ€์ค‘์น˜์˜ ๊ฐ’์„ ๋ฐ”๊พธ๋Š” ํ•™์ŠตยทํŒŒ์ธํŠœ๋‹ยท๋ณ‘ํ•ฉ์€ ํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์›๋ณธ safetensors ๋ฅผ ์•„๋ž˜ ์ ˆ์ฐจ๋กœ ํ˜•์‹ ๋ณ€ํ™˜ + 4bit ์–‘์žํ™”๋งŒ ํ–ˆ์Šต๋‹ˆ๋‹ค.

# llama.cpp @ 7e1e28c (2026-07-28)
python convert_hf_to_gguf.py kanana-2-3b-instruct --outtype bf16 \
  --outfile kanana-2-3b-instruct-BF16.gguf
./build/bin/llama-quantize kanana-2-3b-instruct-BF16.gguf \
  kanana-2-3b-instruct-Q4_K_M.gguf Q4_K_M

4bit ์–‘์žํ™”๋Š” ์›๋ณธ ๋Œ€๋น„ ํ’ˆ์งˆ ์†์‹ค์„ ์ˆ˜๋ฐ˜ํ•ฉ๋‹ˆ๋‹ค. ์›๋ณธ ํ’ˆ์งˆ์ด ํ•„์š”ํ•˜๋ฉด ์œ„ ๋งํฌ์˜ ์›๋ณธ ์ €์žฅ์†Œ๋ฅผ ์“ฐ์„ธ์š”.

์‚ฌ์šฉ๋ฒ•

# llama.cpp
llama-cli -hf DKTechin/kanana:Q4_K_M -c 2048 --jinja \
  -sys "๋‹น์‹ ์€ ์‚ฌ๋‚ด ๋ฉ”์‹ ์ € ๋Œ€ํ™”๋ฅผ ๊ฐ„๊ฒฐํ•˜๊ฒŒ ์š”์•ฝํ•˜๋Š” ๋„์šฐ๋ฏธ์ž…๋‹ˆ๋‹ค." \
  -p "์•„๋ž˜ ๋Œ€ํ™”๋ฅผ 3์ค„๋กœ ์š”์•ฝํ•ด์ค˜: ..."

LM StudioยทJanยทOllama ๋“ฑ GGUF ๋ฅผ ์ฝ๋Š” ๋Ÿฐํƒ€์ž„์—์„œ๋„ ๊ทธ๋Œ€๋กœ ์”๋‹ˆ๋‹ค. ์‹œ์Šคํ…œ ํ”„๋กฌํ”„ํŠธ ์—†์ด ์“ฐ๋ฉด ์š”์•ฝ ๋Œ€์‹  ์ž…๋ ฅ์„ ๋˜ํ’€์ดํ•˜๋Š” ๊ฒฝํ–ฅ์ด ์žˆ์–ด, ์—ญํ• ์„ ์ง€์ •ํ•˜๋Š” ์‹œ์Šคํ…œ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ•จ๊ป˜ ์ฃผ๋Š” ํŽธ์ด ์•ˆ์ „ํ•ฉ๋‹ˆ๋‹ค.

ํ™•์ธํ•œ ๊ฒƒ ยท ํ™•์ธํ•˜์ง€ ๋ชปํ•œ ๊ฒƒ

  • ํ™•์ธ: ๋ฉ”ํƒ€๋ฐ์ดํ„ฐ(archยทtokenizer preยทropeยทchat template), ์‹ค์ œ ํ•œ๊ตญ์–ด ์š”์•ฝ ์ƒ์„ฑ, macOS/Metal ๊ธฐ์ค€ 2048 ํ† ํฐ ์กฐ๊ฑด์—์„œ ํ”„๋กฌํ”„ํŠธ 1,375 t/s ยท ์ƒ์„ฑ 128 t/s.
  • ๋ฏธํ™•์ธ: 32k ์žฅ๋ฌธ ํ’ˆ์งˆ. ์›๋ณธ config.json ์€ rope ๋ฐฐ์ˆ˜๋ฅผ 40 ์œผ๋กœ ์ ์–ด ๋‘์—ˆ์ง€๋งŒ ๊ธธ์ด ๋น„์œจ(32,768 / 4,096)๋กœ ๊ณ„์‚ฐํ•˜๋ฉด 8 ์ด๋ผ ์„œ๋กœ ๋งž์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ์›๋ณธ ํŒŒ์ด์ฌ ๋Ÿฐํƒ€์ž„๊ณผ ๋™์ž‘์„ ๋งž์ถ”๊ธฐ ์œ„ํ•ด ๋ช…์‹œ๊ฐ’ 40 ์„ ๊ทธ๋Œ€๋กœ ๊ธฐ๋กํ–ˆ์œผ๋‹ˆ, ์•„์ฃผ ๊ธด ์ž…๋ ฅ์—์„œ ์ด์ƒํ•˜๋ฉด ์ด ๊ฐ’์„ ๋จผ์ € ์˜์‹ฌํ•˜์„ธ์š”.

๋ผ์ด์„ ์Šค

์ด ํŒŒ์ผ์€ Kanana Open License Agreement ๋ฅผ ๋”ฐ๋ฅด๋Š” ํŒŒ์ƒ๋ฌผ์ž…๋‹ˆ๋‹ค. ์‚ฌ๋ณธ์€ ์ด ์ €์žฅ์†Œ์˜ LICENSE, ๊ณ ์ง€ ๋ฌธ๊ตฌ๋Š” NOTICE ์— ์žˆ์Šต๋‹ˆ๋‹ค.

  • ์‚ฌ์šฉ์ž๋Š” ์›๋ณธ๊ณผ ๋™์ผํ•˜๊ฒŒ ๊ธˆ์ง€๋œ ์‚ฌ์šฉ ์ •์ฑ…(Agreement ยง2.2) ๊ณผ KAKAO ์˜ Guidelines For Responsible AI ๋ฅผ ์ค€์ˆ˜ํ•ด์•ผ ํ•˜๋ฉฐ, ์ด ํŒŒ์ผ์„ ์žฌ๋ฐฐํฌํ•  ๋•Œ์—๋„ ๊ฐ™์€ ์˜๋ฌด๋ฅผ ํ›„์† ์‚ฌ์šฉ์ž์—๊ฒŒ ์•Œ๋ ค์•ผ ํ•ฉ๋‹ˆ๋‹ค.
  • APIยทํด๋ผ์šฐ๋“œ ๋“ฑ์œผ๋กœ ์ œ3์ž์—๊ฒŒ ์ ‘๊ทผ์„ ์ œ๊ณตํ•˜๊ฑฐ๋‚˜ ์žฌํŒ๋งคํ•˜๋ ค๋ฉด KAKAO ์˜ ๋ณ„๋„ ์ƒ์—… ๋ผ์ด์„ ์Šค๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค(Agreement ยง4).
  • ์ด ํŒŒ์ผ์„ ์‚ฌ์šฉํ•˜๋Š” ์›น์‚ฌ์ดํŠธยทUIยท๋ฌธ์„œ์—๋Š” "Powered by Kanana" ๋ฅผ ์•Œ์•„๋ณผ ์ˆ˜ ์žˆ๊ฒŒ ํ‘œ์‹œํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค(Agreement ยง3.1(v)).
Kanana is licensed in accordance with the Kanana Open License Agreement.
Copyright ยฉ KAKAO Corp. All Rights Reserved.
Downloads last month
619
GGUF
Model size
4B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

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

Model tree for DKTechin/kanana

Quantized
(11)
this model