Open Source

Spoiler: while everyone's arguing about Llama vs Qwen, Upstage AI shipped Solar Pro 22B and quietly took the mid-size crown. Korean labs are cooking and the Western LLM bubble is not paying attention.

The Setup

Solar Pro uses Upstage's depth-up-scaling trick — duplicate the middle layers of a strong base, continue pre-training, and you get a much fatter model for cheap. It runs on a single H100 and matches Llama 3 70B on most evals.

huggingface-cli download   upstage/solar-pro-preview-instruct   --local-dir ./solar-pro

# fits on a single 80GB card at bf16, no quant needed

The Money Pattern

The Korean+English bilingual quality is the kicker. Most Western models butcher Korean. Solar Pro switches languages mid-conversation without missing a beat — clutch if you're shipping into APAC markets.

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("upstage/solar-pro-preview-instruct")
model = AutoModelForCausalLM.from_pretrained(
    "upstage/solar-pro-preview-instruct",
    device_map="auto",
    torch_dtype="bfloat16",
)

msgs = [
    {"role": "system", "content": "You are a bilingual KR/EN assistant."},
    {"role": "user", "content": "안녕하세요, can you explain RAG in Korean and English?"},
]
inputs = tok.apply_chat_template(msgs, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(inputs, max_new_tokens=512)[0]))

The Catch

Read the license. The weights are released for research and personal use only — commercial deployment requires a separate agreement with Upstage. For Aidxn Design experiments, fine. For Rebuild Relief production, you're emailing Korea first.

The Verdict

Solar Pro is the model the leaderboard chasers missed. If you can live with the license, it's the strongest 22B you can run on a single GPU. Korean teams are absolutely cooking right now — keep an eye on Upstage and KT.