Solar-Open

Solar Open is Upstage's flagship 102B-parameter large language model, trained entirely from scratch and released under the Upstage Solar License (see LICENSE for details). As a Mixture-of-Experts (MoE) architecture, it delivers enterprise-grade performance in reasoning, instruction-following, and agentic capabilities—all while prioritizing transparency and customization for the open-source community.

Technical Report | Project Page

Highlights

• MoE Architecture (102B / 12B): Built on a Mixture-of-Experts architecture with 102B total / 12B active parameters. This design delivers the knowledge depth of a massive model with the inference speed and cost-efficiency of a much smaller model.
• Massive Training Scale: Pre-trained on 19.7 trillion tokens, ensuring broad knowledge coverage and robust reasoning capabilities across various domains.
• Quantized Version Available: An official INT4 quantized model is provided by NotaAI and available at nota-ai/Solar-Open-100B-NotaMoEQuant-Int4.

Model Overview

• Model Name: Solar Open 100B
• Hugging Face ID: Upstage/Solar-Open-100B
• Architecture: Mixture-of-Experts (MoE)
  • Total Parameters: 102.6B
  • Active Parameters: 12B (per token)
  • Experts: 129 Experts (top 8 among 128 Routed + 1 Shared)
• Pre-training Tokens: 19.7 Trillion
• Context Length: 128k
• Training Hardware: NVIDIA B200 GPUs
• License: Upstage Solar License (See LICENSE)
• Hardware Requirements:
  • Minimum: 4x NVIDIA A100 (80GB)

For more details, please refer to the Solar Open Technical Report.

License

This repository contains both model weights and code, which are licensed under different terms:

1. MODEL WEIGHTS (*.safetensors) Licensed under Upstage Solar License See: https://huggingface.co/upstage/Solar-Open-100B/blob/main/LICENSE

2. CODE (*.py, *.json, *.jinja files) Licensed under Apache License 2.0 See: https://www.apache.org/licenses/LICENSE-2.0

Performance

Korean Benchmarks

English Benchmarks

Inference Quickstart

We recommend using the following generation parameters:

temperature=0.8
top_p=0.95
top_k=50

Transformers

Install the required dependencies:

pip install -U "transformers>=5.0" kernels torch accelerate

Run inference with the following code:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "upstage/Solar-Open-100B"

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

model = AutoModelForCausalLM.from_pretrained(
    pretrained_model_name_or_path=MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

# Prepare input
messages = [{"role": "user", "content": "who are you?"}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt",
)
inputs = inputs.to(model.device)

# Generate response
generated_ids = model.generate(
    **inputs,
    max_new_tokens=4096,
    temperature=0.8,
    top_p=0.95,
    top_k=50,
    do_sample=True,
)
generated_text = tokenizer.decode(generated_ids[0][inputs.input_ids.shape[1] :])
print(generated_text)

vLLM

Option 1: Using Docker (Highly Recommended)

Docker is the recommended deployment method for running Solar-Open-100B.

# For 8 GPUs
docker run --gpus all \
    --ipc=host \
    -p 8000:8000 \
    upstage/vllm-solar-open:latest \
    upstage/Solar-Open-100B \
    --trust-remote-code \
    --enable-auto-tool-choice \
    --tool-call-parser solar_open \
    --reasoning-parser solar_open \
    --logits-processors vllm.model_executor.models.parallel_tool_call_logits_processor:ParallelToolCallLogitsProcessor \
    --logits-processors vllm.model_executor.models.solar_open_logits_processor:Sola