slm / models.yaml
primenumbersam's picture
feat: organize catalog by vendor/size and activate gemma4-e4b and kanana-2-3b
ed0b0be
Raw History Blame Contribute Delete
3.55 kB
vram_budget_gb: 44.0
default_active:
- gemma4-e4b
- kanana-2-3b
max_input_tokens: 8192
max_summary_tokens: 384
catalog:
# ==========================================
# 1. Google (Gemma)
# ==========================================
gemma4-e4b:
name: "Gemma 4 E4B"
repo: "google/gemma-4-E4B-it"
loader: "image-text"
torch_dtype: "bfloat16"
vram_gb: 10.0
description: "구글 온디바이스/엣지 타깃 4B급 경량 모델 (2026-03 공개)"
chat_template_kwargs: {}
gemma4-12b-w4:
name: "Gemma 4 12B (QAT 4-bit)"
repo: "google/gemma-4-12B-it-qat-w4a16-ct"
loader: "image-text"
torch_dtype: "bfloat16"
vram_gb: 10.0
description: "구글 멀티모달 아키텍처 기반 QAT w4a16 압축 모델 (2026-06 공개)"
chat_template_kwargs: {}
# ==========================================
# 2. Alibaba (Qwen)
# ==========================================
qwen2.5-0.5b:
name: "Qwen 2.5 0.5B Instruct"
repo: "Qwen/Qwen2.5-0.5B-Instruct"
loader: "causal-lm"
torch_dtype: "auto"
vram_gb: 1.5
description: "알리바바 0.5B 초경량 고속 모델 (CPU/ZeroGPU 하이브리드)"
chat_template_kwargs: {}
qwen3.5-9b:
name: "Qwen 3.5 9B"
repo: "Qwen/Qwen3.5-9B"
loader: "image-text"
torch_dtype: "bfloat16"
vram_gb: 19.0
description: "알리바바 9B 고성능 경량 모델 (2026-02 공개)"
chat_template_kwargs:
enable_thinking: false
qwen3.8-27b-fp8:
name: "Qwen 3.8 27B (FP8)"
repo: "Qwen/Qwen3.8-27B-FP8"
loader: "image-text"
torch_dtype: "auto"
vram_gb: 30.0
description: "알리바바 27B 하이브리드 아키텍처 공식 FP8 모델 (2026-08 공개)"
chat_template_kwargs:
enable_thinking: false
# ==========================================
# 3. Hugging Face (SmolLM)
# ==========================================
smollm2-1.7b:
name: "SmolLM2 1.7B Instruct"
repo: "HuggingFaceTB/SmolLM2-1.7B-Instruct"
loader: "causal-lm"
torch_dtype: "auto"
vram_gb: 3.5
description: "Hugging Face 공식 1.7B 초경량 고품질 모델 (CPU/ZeroGPU 하이브리드)"
chat_template_kwargs: {}
# ==========================================
# 4. LG AI Research (EXAONE)
# ==========================================
exaone-4.0-1.2b-fp8:
name: "EXAONE 4.0 1.2B (FP8)"
repo: "LGAI-EXAONE/EXAONE-4.0-1.2B-FP8"
loader: "causal-lm"
torch_dtype: "auto"
vram_gb: 2.0
description: "LG AI 초경량 엣지 모델 (2025-07 공개)"
chat_template_kwargs: {}
exaone-4.5-33b-fp8:
name: "EXAONE 4.5 33B (FP8)"
repo: "LGAI-EXAONE/EXAONE-4.5-33B-FP8"
loader: "image-text"
torch_dtype: "auto"
vram_gb: 34.0
description: "LG AI 연구원 한국어/영어 고성능 파운데이션 모델 (2026-04 공개)"
chat_template_kwargs: {}
# ==========================================
# 5. Kakao (Kanana)
# ==========================================
kanana-2-3b:
name: "Kanana 2 3B"
repo: "kakaocorp/kanana-2-3b-instruct"
loader: "causal-lm"
torch_dtype: "bfloat16"
vram_gb: 7.0
description: "카카오 한국어 특화 경량 지시 모델 (2026-07 공개)"
chat_template_kwargs: {}
kanana-2-30b-moe:
name: "Kanana 2 30B MoE"
repo: "kakaocorp/kanana-2-30b-a3b-instruct-2601"
loader: "causal-lm"
torch_dtype: "bfloat16"
vram_gb: 31.0
description: "카카오 30B MoE (활성 3B) 한국어 특화 모델 (2026-01 공개)"
chat_template_kwargs: {}