primenumbersam commited on
Commit
ed0b0be
·
1 Parent(s): 203c07b

feat: organize catalog by vendor/size and activate gemma4-e4b and kanana-2-3b

Browse files
Files changed (1) hide show
  1. models.yaml +55 -61
models.yaml CHANGED
@@ -1,13 +1,24 @@
1
  vram_budget_gb: 44.0
2
  default_active:
3
- - qwen2.5-0.5b
4
- - smollm2-1.7b
5
-
6
 
7
  max_input_tokens: 8192
8
  max_summary_tokens: 384
9
 
10
  catalog:
 
 
 
 
 
 
 
 
 
 
 
 
11
  gemma4-12b-w4:
12
  name: "Gemma 4 12B (QAT 4-bit)"
13
  repo: "google/gemma-4-12B-it-qat-w4a16-ct"
@@ -17,23 +28,27 @@ catalog:
17
  description: "구글 멀티모달 아키텍처 기반 QAT w4a16 압축 모델 (2026-06 공개)"
18
  chat_template_kwargs: {}
19
 
20
- gemma4-31b-w4:
21
- name: "Gemma 4 31B (QAT 4-bit)"
22
- repo: "google/gemma-4-31B-it-qat-w4a16-ct"
23
- loader: "image-text"
24
- torch_dtype: "bfloat16"
25
- vram_gb: 20.0
26
- description: "구글 31B 플래그십 아키텍처 공식 QAT w4a16 압축 모델 (2026-06 공개)"
 
 
 
27
  chat_template_kwargs: {}
28
 
29
- gemma4-e4b:
30
- name: "Gemma 4 E4B"
31
- repo: "google/gemma-4-E4B-it"
32
  loader: "image-text"
33
  torch_dtype: "bfloat16"
34
- vram_gb: 10.0
35
- description: "구글 온디바이스/엣지 타깃 경량 모델 (2026-03 공개)"
36
- chat_template_kwargs: {}
 
37
 
38
  qwen3.8-27b-fp8:
39
  name: "Qwen 3.8 27B (FP8)"
@@ -45,25 +60,21 @@ catalog:
45
  chat_template_kwargs:
46
  enable_thinking: false
47
 
48
- qwen3.5-9b:
49
- name: "Qwen 3.5 9B"
50
- repo: "Qwen/Qwen3.5-9B"
51
- loader: "image-text"
52
- torch_dtype: "bfloat16"
53
- vram_gb: 19.0
54
- description: "알리바바 9B 고성능 경량 모델 (2026-02 공개)"
55
- chat_template_kwargs:
56
- enable_thinking: false
57
-
58
- exaone-4.5-33b-fp8:
59
- name: "EXAONE 4.5 33B (FP8)"
60
- repo: "LGAI-EXAONE/EXAONE-4.5-33B-FP8"
61
- loader: "image-text"
62
  torch_dtype: "auto"
63
- vram_gb: 34.0
64
- description: "LG AI 연구원 한국어/영어 고성능 파운데이션 모델 (2026-04 공개)"
65
  chat_template_kwargs: {}
66
 
 
 
 
67
  exaone-4.0-1.2b-fp8:
68
  name: "EXAONE 4.0 1.2B (FP8)"
69
  repo: "LGAI-EXAONE/EXAONE-4.0-1.2B-FP8"
@@ -73,6 +84,18 @@ catalog:
73
  description: "LG AI 초경량 엣지 모델 (2025-07 공개)"
74
  chat_template_kwargs: {}
75
 
 
 
 
 
 
 
 
 
 
 
 
 
76
  kanana-2-3b:
77
  name: "Kanana 2 3B"
78
  repo: "kakaocorp/kanana-2-3b-instruct"
@@ -90,32 +113,3 @@ catalog:
90
  vram_gb: 31.0
91
  description: "카카오 30B MoE (활성 3B) 한국어 특화 모델 (2026-01 공개)"
92
  chat_template_kwargs: {}
93
-
94
- local-smoke:
95
- name: "Local Smoke (Qwen 0.8B)"
96
- repo: "Qwen/Qwen3.5-0.8B"
97
- loader: "image-text"
98
- torch_dtype: "float32"
99
- vram_gb: 2.0
100
- description: "로컬 CPU 및 파이프라인 무결성 검증용 초경량 모델"
101
- chat_template_kwargs:
102
- enable_thinking: false
103
-
104
- qwen2.5-0.5b:
105
- name: "Qwen 2.5 0.5B Instruct"
106
- repo: "Qwen/Qwen2.5-0.5B-Instruct"
107
- loader: "causal-lm"
108
- torch_dtype: "auto"
109
- vram_gb: 1.5
110
- description: "알리바바 0.5B 초경량 고속 모델 (CPU/ZeroGPU 하이브리드)"
111
- chat_template_kwargs: {}
112
-
113
- smollm2-1.7b:
114
- name: "SmolLM2 1.7B Instruct"
115
- repo: "HuggingFaceTB/SmolLM2-1.7B-Instruct"
116
- loader: "causal-lm"
117
- torch_dtype: "auto"
118
- vram_gb: 3.5
119
- description: "Hugging Face 공식 1.7B 초경량 고품질 요약 모델 (CPU/ZeroGPU 하이브리드)"
120
- chat_template_kwargs: {}
121
-
 
1
  vram_budget_gb: 44.0
2
  default_active:
3
+ - gemma4-e4b
4
+ - kanana-2-3b
 
5
 
6
  max_input_tokens: 8192
7
  max_summary_tokens: 384
8
 
9
  catalog:
10
+ # ==========================================
11
+ # 1. Google (Gemma)
12
+ # ==========================================
13
+ gemma4-e4b:
14
+ name: "Gemma 4 E4B"
15
+ repo: "google/gemma-4-E4B-it"
16
+ loader: "image-text"
17
+ torch_dtype: "bfloat16"
18
+ vram_gb: 10.0
19
+ description: "구글 온디바이스/엣지 타깃 4B급 경량 모델 (2026-03 공개)"
20
+ chat_template_kwargs: {}
21
+
22
  gemma4-12b-w4:
23
  name: "Gemma 4 12B (QAT 4-bit)"
24
  repo: "google/gemma-4-12B-it-qat-w4a16-ct"
 
28
  description: "구글 멀티모달 아키텍처 기반 QAT w4a16 압축 모델 (2026-06 공개)"
29
  chat_template_kwargs: {}
30
 
31
+ # ==========================================
32
+ # 2. Alibaba (Qwen)
33
+ # ==========================================
34
+ qwen2.5-0.5b:
35
+ name: "Qwen 2.5 0.5B Instruct"
36
+ repo: "Qwen/Qwen2.5-0.5B-Instruct"
37
+ loader: "causal-lm"
38
+ torch_dtype: "auto"
39
+ vram_gb: 1.5
40
+ description: "알리바바 0.5B 초경량 고속 모델 (CPU/ZeroGPU 하이브리드)"
41
  chat_template_kwargs: {}
42
 
43
+ qwen3.5-9b:
44
+ name: "Qwen 3.5 9B"
45
+ repo: "Qwen/Qwen3.5-9B"
46
  loader: "image-text"
47
  torch_dtype: "bfloat16"
48
+ vram_gb: 19.0
49
+ description: "알리바바 9B 고성능 경량 모델 (2026-02 공개)"
50
+ chat_template_kwargs:
51
+ enable_thinking: false
52
 
53
  qwen3.8-27b-fp8:
54
  name: "Qwen 3.8 27B (FP8)"
 
60
  chat_template_kwargs:
61
  enable_thinking: false
62
 
63
+ # ==========================================
64
+ # 3. Hugging Face (SmolLM)
65
+ # ==========================================
66
+ smollm2-1.7b:
67
+ name: "SmolLM2 1.7B Instruct"
68
+ repo: "HuggingFaceTB/SmolLM2-1.7B-Instruct"
69
+ loader: "causal-lm"
 
 
 
 
 
 
 
70
  torch_dtype: "auto"
71
+ vram_gb: 3.5
72
+ description: "Hugging Face 공식 1.7B 초경량 고품질 모델 (CPU/ZeroGPU 하이브리드)"
73
  chat_template_kwargs: {}
74
 
75
+ # ==========================================
76
+ # 4. LG AI Research (EXAONE)
77
+ # ==========================================
78
  exaone-4.0-1.2b-fp8:
79
  name: "EXAONE 4.0 1.2B (FP8)"
80
  repo: "LGAI-EXAONE/EXAONE-4.0-1.2B-FP8"
 
84
  description: "LG AI 초경량 엣지 모델 (2025-07 공개)"
85
  chat_template_kwargs: {}
86
 
87
+ exaone-4.5-33b-fp8:
88
+ name: "EXAONE 4.5 33B (FP8)"
89
+ repo: "LGAI-EXAONE/EXAONE-4.5-33B-FP8"
90
+ loader: "image-text"
91
+ torch_dtype: "auto"
92
+ vram_gb: 34.0
93
+ description: "LG AI 연구원 한국어/영어 고성능 파운데이션 모델 (2026-04 공개)"
94
+ chat_template_kwargs: {}
95
+
96
+ # ==========================================
97
+ # 5. Kakao (Kanana)
98
+ # ==========================================
99
  kanana-2-3b:
100
  name: "Kanana 2 3B"
101
  repo: "kakaocorp/kanana-2-3b-instruct"
 
113
  vram_gb: 31.0
114
  description: "카카오 30B MoE (활성 3B) 한국어 특화 모델 (2026-01 공개)"
115
  chat_template_kwargs: {}