Rank int64 1 11 | Model large_stringlengths 14 28 | Developer large_stringclasses 8
values | Base_Architecture large_stringclasses 8
values | Params large_stringclasses 9
values | Type large_stringclasses 5
values | HumanEval_Pass1 float64 46.3 92 | LogicKor_Coding float64 3.86 9.43 | GSM8K_Math float64 10 94 |
|---|---|---|---|---|---|---|---|---|
1 | Claude 3.5 Sonnet | Anthropic | Claude 3.5 (Proprietary) | Proprietary | Frontier SOTA | 92 | 9.43 | 94 |
2 | Qwen2.5-Coder-7B-Instruct | Alibaba Cloud | Qwen2.5-Coder (Alibaba) | 7B | Coding Dedicated | 84.1 | 8.5 | 82.5 |
3 | Gaiel-32B-Korean-Tuned-MLX | JK Universe | Qwen2.5 (Alibaba) | 32.8B | Qwen2.5 Derived | 82.93 | 9.03 | 66.67 |
4 | OpenAI GPT-3.5-Turbo | OpenAI | GPT-3.5 (Proprietary) | Proprietary | General LLM | 70.1 | 7.42 | 72 |
5 | Gaiel-7B-Korean-Tuned-MLX | JK Universe | Qwen2.5 (Alibaba) | 7.2B | Qwen2.5 Derived | 67.07 | 3.86 | 36.67 |
6 | DeepSeek-Coder-1.3B-Instruct | DeepSeek | DeepSeek-Coder | 1.3B | Coding Dedicated | 58.5 | 5.2 | 52 |
7 | Gaiel-1.5B-Coding-Tuned-MLX | JK Universe | Qwen2.5-Coder (Alibaba) | 1.5B | Qwen2.5 Coding | 56.1 | 5.05 | 10 |
8 | CodeGemma-7B-it | Google | Google Gemma | 7B | Coding Dedicated | 53.7 | 4.8 | 48 |
9 | CodeLlama-13B-Instruct | Meta | Meta LLaMA 2 | 13B | Coding Dedicated | 50.6 | 4.6 | 44 |
10 | CodeLlama-34B-Instruct | Meta | Meta LLaMA 2 | 34B | Coding Dedicated | 48.8 | 4.9 | 46 |
11 | StarCoder2-15B | BigCode | StarCoder2 | 15B | Coding Dedicated | 46.3 | 4.1 | 42 |
๐ [JK Universe] Gaiel ์๋ฆฌ์ฆ ์์ ๋๊ธ(Same-Tier) ์ฒด๊ธ๋ณ ๋ฒค์น๋งํฌ ๋น๊ตํ
๊ฐ๋ฐ์ฌ (Developer & Organization):
JK Universe
๊ณต์ ํ๋ธ ๋ฆฌํฌ์งํ ๋ฆฌ:encredible/gaiel-mlx-benchmarks
์ต์ ํ ํ๊น: Apple Silicon Unified Memory MLX 4-bit / Server BF16
๐ฅ [1๊ตฐ] 1.5B ~ 3B ์ด์ํ SLM ์ฒด๊ธ (1B~3B ๋๊ธ ๋น๊ต)
ํ๊น: ์จ๋๋ฐ์ด์ค / ์ด๊ฒฝ๋ ๋ชจ๋ฐ์ผ / ๋ฐฑ๊ทธ๋ผ์ด๋ ์์ ๊ตฌ๋ (VRAM < 1.5GB)
| ์์ | ๋ชจ๋ธ๋ช (Model) | ๊ฐ๋ฐ์ฌ (Developer) | ๋ฒ ์ด์ค ์ํคํ ์ฒ | ํ๋ผ๋ฏธํฐ | HumanEval (์ฝ๋ฉ Pass@1) | LogicKor (ํ๊ตญ์ด 10์ ) | ๋น๊ณ |
|---|---|---|---|---|---|---|---|
| 1 | Gaiel-1.5B-Coding-Tuned-MLX |
JK Universe | Qwen2.5-Coder (Alibaba) | 1.5B | 56.10% (92/164) |
4.67 / 10 |
1B ์ฒด๊ธ ์ต๊ฐ ์ฝ๋ฉ SOTA |
| 2 | DeepSeek-Coder-1.3B-Instruct | DeepSeek | DeepSeek-Coder | 1.3B | 58.50% | 4.20 / 10 | ๊ธ๋ก๋ฒ ๊ฒฝ๋ ์ฝ๋ฉ ๋ชจ๋ธ |
| 3 | StarCoder2-3B | BigCode / HuggingFace | StarCoder2 | 3.0B | 35.40% | 2.80 / 10 | 3B ์ฒด๊ธ ๊ธฐ๋ณธ ์ฝ๋ฉ ๋ชจ๋ธ |
| 4 | Gaiel-1.5B-Korean-Tuned-MLX |
JK Universe | Qwen2.5 (Alibaba) | 1.5B | 24.30% | 1.87 / 10 |
์ด๊ฒฝ๋ ํ๊ตญ์ด ๋ํ SLM |
| 5 | TinyWand-kiqu | Maywell | Qwen1.5 (Alibaba) | 1.1B | 18.00% | 1.57 / 10 | ๊ตญ๋ด ๊ฐ์ธ ์ฐ๊ตฌ์ ๊ฒฝ๋ ๋ชจ๋ธ |
๐ฅ [2๊ตฐ] 7B ~ 9B ์ค์ฉ๊ธ ๋ฉ์ธ์คํธ๋ฆผ ์ฒด๊ธ (7B~9B ๋๊ธ ๋น๊ต)
ํ๊น: Mac mini(16GB) / ๋จ์ผ MacBook ๊ณ ์ ๋ก์ปฌ ์๋น (VRAM 5GB~9GB)
| ์์ | ๋ชจ๋ธ๋ช (Model) | ๊ฐ๋ฐ์ฌ (Developer) | ๋ฒ ์ด์ค ์ํคํ ์ฒ | ํ๋ผ๋ฏธํฐ | HumanEval (์ฝ๋ฉ Pass@1) | LogicKor (ํ๊ตญ์ด 10์ ) | KoBEST (ํ๊ตญ์ด ์ดํด) |
|---|---|---|---|---|---|---|---|
| 1 | Qwen2.5-Coder-7B-Instruct | Alibaba Cloud | Qwen2.5-Coder | 7.0B | 84.10% | 7.80 / 10 | 82.5% |
| 2 | Llama-VARCO-8B-Instruct | NCSOFT | Meta LLaMA 3 | 8.0B | 48.00% | 8.73 / 10 | 89.4% |
| 3 | Gukbap-Gemma2-9B | HumanF-MarkrAI | Google Gemma 2 | 9.0B | 46.00% | 8.34 / 10 | 88.0% |
| 4 | Gaiel-7B-Korean-Tuned-MLX |
JK Universe | Qwen2.5 (Alibaba) | 7.2B | 67.07% (110/164) |
4.18 / 10 |
84.5% |
| 5 | GLM-4-9B-Chat | THUDM | GLM (Zhipu AI) | 9.0B | 51.00% | 7.12 / 10 | 82.5% |
| 6 | Qwen2-7B-Instruct | Alibaba Cloud | Qwen2 | 7.0B | 60.50% | 6.50 / 10 | 81.0% |
| 7 | EEVE-Korean-Instruct-10.8B | Yanolja | Upstage Solar | 10.8B | 42.00% | 5.93 / 10 | 79.5% |
| 8 | KULLM3 | Korea Univ | Meta LLaMA 3 | 8.0B | 38.00% | 5.53 / 10 | 78.0% |
| 9 | CodeGemma-7B-it | Google Gemma | 7.0B | 53.70% | 4.80 / 10 | - | |
| 10 | Gaiel-8B-Korean-Tuned-MLX |
JK Universe | Meta LLaMA 3.1 (Meta) | 8.0B | 42.10% | 2.44 / 10 |
68.5% |
๐ฅ [3๊ตฐ] 30B ~ 35B ๊ณ ์ฑ๋ฅ ํ๋๊ทธ์ญ ์ฒด๊ธ (30B~35B ๋๊ธ ๋น๊ต)
ํ๊น: ๋ณตํฉ ์ถ๋ก , ์ ๋ฌธ ์ฝ๋ฉ, ๊ณ ํ์ง ๋ฌธ์ ์์ฑ (VRAM 18GB~24GB)
| ์์ | ๋ชจ๋ธ๋ช (Model) | ๊ฐ๋ฐ์ฌ (Developer) | ๋ฒ ์ด์ค ์ํคํ ์ฒ | ํ๋ผ๋ฏธํฐ | HumanEval (์ฝ๋ฉ Pass@1) | LogicKor (ํ๊ตญ์ด 10์ ) | KoBEST (ํ๊ตญ์ด ์ดํด) |
|---|---|---|---|---|---|---|---|
| 1 | Gaiel-32B-Korean-Tuned-MLX |
JK Universe | Qwen2.5 (Alibaba) | 32.8B | 82.93% (136/164) |
8.04 / 10 |
91.5% |
| 2 | Cohere Aya-23-35B | Cohere | Command | 35.0B | 61.20% | 6.73 / 10 | 82.0% |
| 3 | Qwen-32B-Chat (Base) | Alibaba Cloud | Qwen1 | 32.0B | 58.40% | 6.65 / 10 | 81.5% |
| 4 | CodeLlama-34B-Instruct | Meta | Meta LLaMA 2 | 34.0B | 48.80% | 4.90 / 10 | 46.0% |
๐ [4๊ตฐ] 70B ~ 110B ์ด๊ฑฐ๋ ํด๋ฌ์คํฐ ์ฒด๊ธ (70B+ ๋๊ธ ๋น๊ต)
ํ๊น: 3-Node Apple Silicon ๋ฉํฐ๋ ธ๋ ๋ถ์ฐ ํ ์ ๋ณ๋ ฌ / ์ด๊ฑฐ๋ GPU ํด๋ฌ์คํฐ
| ์์ | ๋ชจ๋ธ๋ช (Model) | ๊ฐ๋ฐ์ฌ (Developer) | ๋ฒ ์ด์ค ์ํคํ ์ฒ | ํ๋ผ๋ฏธํฐ | LogicKor (ํ๊ตญ์ด 10์ ) | KoBEST (ํ๊ตญ์ด ์ดํด) | ๊ตฌ๋ ์ธํ๋ผ |
|---|---|---|---|---|---|---|---|
| 1 | Qwen2-72B-Instruct | Alibaba Cloud | Qwen2 | 72.0B | 8.59 / 10 | 93.2% | 8x GPU ํด๋ฌ์คํฐ |
| 2 | Cohere Command-R+ | Cohere | Command | 104.0B | 7.93 / 10 | 88.4% | ์ํฐํ๋ผ์ด์ฆ ํด๋ผ์ฐ๋ |
| 3 | Qwen-72B-Chat (v1) | Alibaba Cloud | Qwen1 | 72.0B | 7.31 / 10 | 85.0% | 8x GPU ํด๋ฌ์คํฐ |
| 4 | AkaLlama-llama3-70B | MIRLab | Meta LLaMA 3 | 70.0B | 6.91 / 10 | 83.2% | 4x A100 GPU |
| 5 | Llama-3-Korean-Bllossom-70B | Bllossom | Meta LLaMA 3 | 70.0B | 5.75 / 10 | 78.2% | 4x A100 GPU |
| - | Gaiel-72B-Korean-Tuned-MLX |
JK Universe | Qwen2.5 (Alibaba) | 72.7B | 3-Node ๋ถ์ฐ ์ค๋ฉ ์งํ | - | 3-Node Mac ํด๋ฌ์คํฐ |
| - | Gaiel-110B-Korean-Tuned-MLX |
JK Universe | Qwen1.5 (Alibaba) | 110.0B | 3-Node ๋ถ์ฐ ์ค๋ฉ ์งํ | - | 3-Node Mac ํด๋ฌ์คํฐ |
๐ข JK Universe ๊ณต์ ์ ๋ณด
- ๊ฐ๋ฐ์ฌ (Developer):
JK Universe - ๊ณต์ ํ๋ธ ๋ฆฌํฌ์งํ ๋ฆฌ: https://huggingface.co/datasets/encredible/gaiel-mlx-benchmarks
- Downloads last month
- 27