Solar LLM

Upstage's Solar LLMs — Korea's frontier family, now Solar Pro 3

LLMs & Chat Open Source Has API Open Source
Researched · Published · Reviewed
RECATOOLS Score
7 / 10
Capability
7
Value for money
8
Ease of use
5
ASEAN readiness
7
API quality
7
Founded
2020
HQ
Seoul, South Korea
Users
200k+ downloads
Launched
Dec 2023
Developer
Upstage AI

Overview

Upstage's Solar family runs from the Apache-2.0 Solar 10.7B (2023) to Solar Pro 3 (January 2026), a 128K-context MoE served via API at $0.15/$0.60 per million tokens. It offers the strongest Korean-language performance of any commercial model family.

Advertisement

Pricing

Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.

Free
Free
Free to download

Use cases

Korean-English bilingual applications requiring strong performance in both languages Efficient deployment in APAC markets with Korean language requirements Research into depth upscaling as an alternative to full pre-training
Advertisement

ASEAN Perspective

Solar LLM in Southeast Asia

ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).

RECATOOLS Verdict

Solar is the rare 2023 open-weights story that grew into a real product line. The Apache-2.0 Solar 10.7B and its depth up-scaling trick topped the Hugging Face leaderboard in December 2023; the commercial line then kept shipping — Solar Pro 2 (31B, July 2025) scored above GPT-4.1 on Artificial Analysis's Intelligence Index, the first Korean-built model rated frontier-class, and Solar Pro 3 (January 2026) is a 128K-context mixture-of-experts model at $0.15/$0.60 per million tokens with a free window until March 2026. Korean-language performance is the standout, and Upstage's document-AI products give it real APAC enterprise traction. Caveats: a smaller ecosystem than the US majors, and the old 10.7B open weights are far behind current open models — the API line is where the capability lives.

Independent AI-assisted assessment by RECATOOLS.

What people say

Fifteen cents per million input tokens, 60 for output — that's Solar Pro 3 pricing at its January 2026 launch, and Upstage is giving it away free until March 2 to pull developers in. That's the current state of a line most directories still describe as 'the 10.7B model that topped the Hugging Face leaderboard in December 2023.' Both things are true; only one matters now.

The 2023 story earned its keep. Depth up-scaling — splicing extra transformer layers into a pretrained base and continuing training — put Solar 10.7B briefly ahead of Mistral 7B and Llama 2 13B, and those Apache-2.0 weights are still on Hugging Face. But Upstage kept shipping: Solar Pro (22B, late 2024), then Solar Pro 2 (31B, July 2025), which Artificial Analysis scored at 58 on its Intelligence Index — above GPT-4.1, and the first Korean-built model it rated frontier-class. Solar Pro 3 (January 2026) moved to a mixture-of-experts design with a 128K context window and selectable chat and reasoning modes.

The pitch that holds up under scrutiny is Korean. On Ko-MMLU, Hae-Rae and Arena-Hard-style benchmarks the Pro models match or beat the US giants, and Upstage pairs the LLM with a document-parse/OCR product line that does real enterprise business in Korea and Japan. For Southeast Asian teams the case is thinner — this is a Korea-first, APAC-second stack — but as a cost-efficient GPT-4-class API from outside the US-China duopoly, it's worth benchmarking. Self-hosting the old 10.7B in 2026 is nostalgia; the capability lives in the API line.

Summary of public user & expert reviews, compiled by RECATOOLS.

Notable facts

  • Solar 10.7B was trained by a South Korean company and outperformed all European and American models of similar size, demonstrating APAC's growing contributions to frontier AI.
  • The Depth Upscaling technique used for Solar allows models to be built more efficiently by combining and extending existing pre-trained architectures.
  • Solar was the first non-Western model to achieve #1 on the HuggingFace Open LLM Leaderboard.

Frequently asked questions

Is Solar free?
Yes. Model weights are free under the Apache 2.0 licence.
What makes Solar's architecture different?
Depth Upscaling extends pre-trained model layers rather than training from scratch, making training more efficient.
Does Solar support Korean?
Yes. Strong Korean and English bilingual capability is a key feature.
How does Solar compare to Mistral 7B?
Solar 10.7B outperforms Mistral 7B on most benchmarks while being somewhat larger.
Can Solar be used commercially?
Yes. Apache 2.0 licence permits commercial use.

About this listing

Researched on
Published on
Last reviewed

This entry was compiled from publicly available data including Solar LLM's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Solar LLM unless explicitly stated.

Data accuracy

Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.

For the latest details, please refer to Solar LLM directly →

Spotted something out of date? Suggest an update →

Advertisement