APMIC CaiGunn Model

CaiGunn 34B CaiGunn 34Bx2 (Coming soon)
Training architecture APMIC APMIC Brainformers + Llama Mamba + Transformer + MoDE
Computing architecture NeMo Framework ANF
Hardware requirements
  • Recommended GPU VRAM needs to exceed 70G
  • Compatible with GPUs of A100, H100, H200, B100 and above
  • Recommended GPU VRAM needs to exceed 80G~140GB
  • Compatible with GPUs of A100, H100, H200, B100 and above
Features
  • checkSupport English and Chinese
  • checkCan run locally
  • check4K Context window
  • checkCan be fine-tuned with the CaiGunn platform
  • checkMMLU score higher than GPT 3.5
  • checkSupports English and Chinese
  • checkCan run locally
  • check8K Context window
  • checkCan be fine-tuned with the CaiGunn platform

Cost halved, with no compromise on performance

  • check After completing Imitation Learning for APMIC, the inference cost can be reduced by over 50%.
  • check Outperforms well-known models in Chinese and English according to specific evaluation criteria.
  • check Supports domain-specific fine-tuning via the CaiGunn platform.

MMLU Rankings (English)

Company Name Model Score
OpenAI GPT-4Not deployable 86.5
APMIC CaiGunn 75.7
Google Gemini ProNot deployable 71.8
Mistral AI Mixtral-8x7B 71.4
OpenAI GPT-3.5Not deployable 70.0
Meta LLaMA 65B 68.9
Google Gemma 7B 64.6

TMMLU+ Rankings (Traditional Chinese)

Company Name Model Score
OpenAI GPT-4Not deployable 60.40
APMIC CaiGunn-zh 55.20
Media Tek Breeze-7B 40.35
Mistral AI Mixtral-8x7B 36.93
NTU Taiwan-LLM-13B 21.36
Innolux Bailong-instruct-7B 6.80

Capable of Handling 21 Times the Data of OpenAI

Input Data Volume Capable of Processing at the Same Cost

Company Name Model Tokens
APMIC CaiGunn 21.4x
Google Gemini 12.0x
OpenAI GPT3.5-Turbo 1x

Output Data Volume Capable of Processing at the Same Cost

Company Name Model Tokens
APMIC CaiGunn 8.7x
Google Gemini 5.3x
OpenAI GPT3.5-Turbo 1x

Flexible Free Deployment

Instant Cloud Usage

Instantly use well-known large language models through CaiGunn. The platform includes built-in automatic text preprocessing, RAG, image-text output, version control, model testing, preview, deployment, and more. It also supports custom model training, fine-tuning, and inference in the cloud, allowing you to have your own models in the cloud.

Enterprise On-Premises Usage

CaiGunn's enterprise version offers powerful features, supporting not only NVIDIA DGX and HGX hardware but also hybrid and private cloud architectures, including AWS, Azure, Google, Oracle Cloud, DGX Cloud, and other cloud providers, ensuring data privacy and protection.

Developer Zone

CaiGunn's models are currently deployed in Taiwan. We provide users with model usage through the CaiGunn API, enabling not only No-Code usage but also secondary development.

Low Inference Cost

After fine-tuning, models trained in specific domains can provide lower inference costs.

Support for Application Scenarios

Language models are just the core. We can support more application systems launched by APMIC, such as customer service, knowledge management, contract recognition, etc.

Data Confidentiality

With NVIDIA H100 or higher-level Confidential Computing technology, data can be kept confidential from training to deployment.

View Developer Documentation Start Free

Create your own GPT now