Upstage

  Executive Briefing

Upstage is a South Korean full-stack enterprise AI company founded in October 2020, headquartered in Yongin, South Korea, with offices in San Jose and Tokyo. The company builds the Solar family of large language models and a Document AI suite targeting regulated industries — insurance, finance, healthcare, and legal — where unstructured document processing is the primary bottleneck to AI adoption. In April 2026, Upstage achieved a valuation exceeding KRW 1 trillion (approximately $680 million reported), becoming the first South Korean generative AI startup to reach unicorn status.¹

The company is led by its three co-founders, all formerly senior executives at Naver's Clova AI division: Sung Kim (CEO), who built and ran Clova's research organization from three researchers to 150 before returning to an academic chair at HKUST and then launching Upstage; Hwalsuk Lee (CTO, also known as Stan Lee), who led Clova's Visual AI and OCR team; and Lucy Park (CSO), former head of the Papago machine translation modeling team and creator of KoNLPy, the most widely used open-source Korean NLP library.² Their combined expertise in Korean language AI, computer vision, and enterprise deployment shaped Upstage's dual-product strategy from the outset.

Upstage occupies a distinctive position in the global AI landscape: it competes not as a pure research institution or model API reseller but as an end-to-end enterprise AI provider building its own weights, training infrastructure, APIs, and customer-facing products. Its Solar Pro 3 model — a mixture-of-experts architecture with 102B total but only 12B active parameters per forward pass — achieved a Tau2-all agentic benchmark score of 72.3 as of March 2026, versus a score of 36.0 for its predecessor, underpinning the company's claim that efficiency-first architecture can match frontier-scale models at dramatically lower cost. The company is simultaneously pursuing a KOSPI IPO (targeted for the second half of 2026 or first half of 2027), a planned acquisition of Daum (South Korea's second-largest search portal) from Kakao, and a role in Korea's government-backed Dokpa-mo Sovereign AI Foundation Model Project.

  At a Glance

  Origins & Founding

Upstage was founded in October 2020 by Sung Kim, Hwalsuk Lee, and Lucy Park, all departing Naver's Clova AI division within the same year.² Kim had been Clova's founding head, growing the organization from three to 150 researchers between 2017 and 2020; before Naver he held a professorship at Hong Kong University of Science and Technology (HKUST), where he earned four ACM Sigsoft Distinguished Paper awards for work in automated bug prediction and code generation, and earlier co-founded Nara Vision in 1995 to build KkebiMail, one of Korea's first webmail services. Lee had led Clova's Visual AI team, specializing in OCR and document-understanding computer vision. Park had headed the Papago translation modeling team and authored KoNLPy, the open-source Korean NLP toolkit now foundational to academic and commercial Korean language research.

The founding thesis was that the primary barrier to enterprise AI adoption was not a shortage of data or IT resources but a shortage of organizations capable of bridging domain data to working AI systems.² The name "Upstage" encodes this intent: to bring companies "up" to AI's potential. Beyond its Naver heritage, the founding team recruited talent from Meta, Amazon, Nvidia, eBay, Google, Apple, and Kakao, building a team with an unusually broad combination of model research and product delivery experience for a seed-stage company. Within the first year Upstage fielded more than 100 enterprise requests across finance, education, IT, and manufacturing, signaling that the enterprise-first positioning had immediate demand.

  History & Timeline

    2020–2022: Seed phase and competition victories

Upstage launched without an initial public product, focusing instead on enterprise engagements and competitive benchmarks as proof of technical credibility. Within 18 months of founding the team won more than 10 Kaggle gold medals — at the time making Upstage the first Korean company to accumulate double-digit gold medals — establishing its standing in applied ML competitions.² In 2021 the company raised a Series A of KRW 31.6 billion (approximately $23 million), backed by SBVA (then SoftBank Ventures Asia), Primer Sazze Partners, Company K Partners, and Premier Partners.

    2023: Leaderboard dominance and the Solar 10.7B launch

In August 2023 Upstage's 70B language model reached the top of the Hugging Face Open LLM Leaderboard, outperforming GPT-3.5 and making Upstage a recognized name outside Korea for the first time.³ The more consequential milestone arrived in December 2023 with the release of Solar 10.7B, a 10.7-billion-parameter model that again topped the leaderboard despite being a fraction of the size of competing models.⁴ Solar 10.7B was built using a proprietary technique called Depth Up-Scaling (DUS) — described in arXiv preprint 2312.15166 — which scales a model's effective capacity through depthwise layer addition and continued pretraining rather than the mixture-of-experts complexity then common at larger scales. The model was released open-source under Apache 2.0, accelerating adoption and establishing Upstage's reputation as an efficiency-first model lab.

    2024: Series B, AWS partnership, and enterprise products

January 2024 brought a further benchmark achievement: Upstage's 13B LLM surpassed Microsoft ToRA and ChatGPT on mathematical reasoning benchmarks.³ In February 2024 the company launched the Solar API Beta (free access), signed a partnership with LG Electronics to embed Solar on LG Gram laptops as an on-device LLM, and partnered with Polaris Office for on-device AI document software. By March 2024 Solar Mini was available on AWS SageMaker, and Upstage received the 2023 AI Korea Grand Prize (Minister of Science and ICT Award).

The financing landmark of the year was the Series B, announced 16 April 2024: $72 million led by SK Networks and KT, with participation from Korea Development Bank, Shinhan Venture Investment, Hana Ventures, Mirae Asset Venture Investment, and the Industrial Bank of Korea, alongside returning investors SBVA, Primer Sazze Partners, Company K Partners, and Premier Partners.⁵ Cumulative investment crossed $100 million, making Upstage the most-funded South Korean AI software company at that point. To accompany the raise, Upstage opened its San Jose, CA office to lead US market expansion. In June 2024 the company signed a multi-year Strategic Collaboration Agreement with AWS naming it Upstage's preferred cloud provider, with Solar models listed on Amazon Bedrock, SageMaker JumpStart, and the AWS Marketplace.

September 2024 saw the preview release of Solar Pro (22B parameters), with general availability in December 2024 on AWS Bedrock and the AWS Marketplace, delivering a reported 50% benchmark improvement over Solar Mini.⁶

    2025: International expansion, Solar Pro 2, and the Series B Bridge

The first half of 2025 brought Upstage into the Japanese market. In March 2025 Upstage and Tokyo-based Karakuri Inc. launched Syn Pro, a Japanese-enterprise LLM with under 14 billion parameters, which immediately ranked first on the Nejumi Leaderboard for Japanese LLMs under 14B. In April 2025 Upstage was named to the CB Insights 100 Most Innovative AI Startups list. A June 2025 partnership with FuriosaAI extended Solar models to FuriosaAI's next-generation AI chips.

The defining model release of 2025 was Solar Pro 2, launched 10 July 2025: 31B parameters with a dual Chat Mode / Reasoning Mode incorporating chain-of-thought for complex tasks, agentic multi-step task execution, and benchmark performance that matched GPT-4o, DeepSeek R1, and Mistral Small 3.2 — models with substantially larger or differently scaled architectures.⁷ Upstage positioned Solar Pro 2 as South Korea's first frontier-class LLM.

In August 2025 Upstage raised a $45M Series B Bridge led by Korea Development Bank, Amazon (AWS), and AMD, bringing cumulative reported capital to approximately $157 million and cementing AWS as Upstage's preferred cloud partner across the Asia-Pacific region and the United States.⁸ In November 2025 Japan's Ministry of Economy, Trade and Industry (METI) certified Syn Pro as a domestic foundation model. In December 2025 Upstage hired KB Securities and Mirae Asset Securities as IPO underwriters, targeting a KOSPI listing in the second half of 2026 or first half of 2027.⁹

On 17 December 2025 Upstage released Solar Open 100B: a mixture-of-experts model with 102B total parameters and approximately 12B active parameters per forward pass (128 routed experts plus one shared expert), pre-trained on 19.7 trillion tokens, and released as open-weight under the Solar-Apache License 2.0.¹⁰

    2026: Unicorn status, Solar Pro 3, and Daum acquisition

January 2026 opened with controversy: Saionic AI CEO Ko Seok-hyun published a GitHub analysis claiming Solar Open 100B was 96.8% identical to Chinese model GLM 4.5 Air at the LayerNorm layer, alleging the model was not domestically built — a serious allegation given Upstage's participation in Korea's Dokpa-mo sovereign AI project and its pending IPO (see Notable Events). Despite the dispute, Upstage passed the first phase of the Dokpa-mo evaluation in January 2026 alongside LG AI Research and SK Telecom.

26 January 2026 brought the release of Solar Pro 3: a MoE architecture with 128K context, trained using the SnapPO reinforcement learning framework, delivering a reported 52% improvement in instruction-following and 30% improvement in complex reasoning over its predecessor.¹¹ An updated release in March 2026 pushed the model's Tau2-all agentic benchmark score to 72.3, compared to 36.0 for Solar Pro 2 — a 2× improvement in multi-step agentic task performance.¹²

On 29 January 2026 Upstage announced an agreement to acquire Daum, South Korea's second-largest search portal, from Kakao in an all-stock deal, formally signing the agreement on 7 May 2026.¹³ In March 2026 Upstage deepened its AMD partnership, announcing plans to deploy MI355 GPUs for sovereign AI infrastructure in Korea, with AMD reported to be in talks to sell 10,000 MI355 chips to the company.¹⁴

On 15 April 2026 Upstage announced the first tranche of a Series C round: KRW 180 billion (approximately $126 million) led by Silicon Valley firm Sazze Partners, valuing the company at over KRW 1 trillion — approximately $680 million reported — and making Upstage the first South Korean generative AI startup to achieve unicorn status.¹

  Mission, Philosophy & Research Agenda

Upstage's stated mission is "Making AI beneficial" — a commitment to ensuring that the practical gains of AI are accessible to enterprises and, through them, to the societies they operate in. The company's internal operating philosophy targets what it calls "AGI for work": purpose-built AI that automates the document-heavy, reasoning-intensive workflows that dominate regulated industries, rather than general consumer-facing AI.² This distinction shapes every product decision: Solar models are evaluated on enterprise task accuracy, instruction-following, and agentic capability rather than on broad consumer utility or raw benchmark supremacy.

A core philosophical commitment is to domestically developed AI — homegrown model weights built from first principles rather than derivatives of foreign base models. This positioning is both a business strategy (enabling participation in Korea's sovereign AI program) and a research imperative: Upstage's primary architectural contribution, Depth Up-Scaling (DUS), arose from the ambition to achieve frontier-class performance without the compute scale available only to hyperscalers.⁴ The SnapPO reinforcement learning framework introduced with Solar Pro 3 continues this tradition of methodological innovation aimed at efficiency. The document intelligence product line — combining OCR accuracy above 95% with structured output in HTML and Markdown — reflects the research team's long heritage in visual AI and its conviction that unstructured documents are the specific bottleneck blocking enterprise AI in insurance, finance, healthcare, and legal sectors.

  Research & Publications

Upstage's most significant research contribution to the broader field is the Depth Up-Scaling (DUS) methodology, introduced in the paper "SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling" (arXiv:2312.15166, December 2023).⁴ Authored by a team of 18 researchers including Dahyun Kim, Chanjun Park, and Sunghun Kim, the paper describes a technique that expands a base model's depth through layer duplication and continued pretraining — achieving high-quality performance without the architectural complexity and routing overhead of mixture-of-experts designs. The resulting Solar 10.7B model topped the Hugging Face Open LLM Leaderboard on release under Apache 2.0, making the technique immediately reproducible.

Upstage's research themes cluster around five areas: efficient LLM scaling (high performance per active parameter); Korean and Japanese multilingual NLP; OCR and document intelligence; agentic reasoning and tool use; and reinforcement learning for instruction-following — most recently the SnapPO framework that trained Solar Pro 3.¹¹ In competition settings, the team won four categories at the ICDAR Robust Reading Competition, the world's premier OCR benchmark, and accumulated more than 10 Kaggle gold medals — the first Korean company to reach double-digit gold medals.

Open-source releases include Solar 10.7B and Solar 10.7B-Instruct (Apache 2.0) and Solar Open 100B (Solar-Apache License 2.0). KoNLPy, the Korean NLP library authored by co-founder Lucy Park before Upstage's founding, remains a widely cited foundational resource for Korean language research.

  Models & Products

Upstage ships models via the Solar API (also called the Upstage Console), through cloud marketplace listings on AWS and Oracle, and via direct enterprise deployment. The product surface spans two pillars: frontier LLMs in the Solar family and a Document AI suite centered on Document Parse.

Solar LLM family:

• Solar 10.7B — Released Dec 2023; 10.7B parameters via DUS; open-weight under Apache 2.0; topped Hugging Face Open LLM Leaderboard on release.
• Solar 10.7B-Instruct — Instruction-tuned variant; outperformed Mixtral-8x7B-Instruct on release benchmarks.
• Solar Mini — Compact, fast variant; reported 2.5× faster than GPT-3.5 at comparable performance; available on AWS SageMaker from Mar 2024.
• Solar Pro — 22B parameters; GA Dec 2024 on AWS Bedrock, SageMaker JumpStart, and AWS Marketplace; 50% benchmark improvement over Solar Mini.⁶
• Solar Pro 2 — 31B parameters; Jul 2025; dual Chat and Reasoning modes; agentic multi-step execution; matched GPT-4o and DeepSeek R1 on reported benchmarks; South Korea's first frontier-class LLM.⁷
• Solar Open 100B — 102B total / ~12B active MoE; Dec 2025; pre-trained on 19.7T tokens; open-weight under Solar-Apache License 2.0.¹⁰
• Solar Pro 3 — MoE; 128K context; SnapPO RL training; Tau2-all agentic score 72.3 (Mar 2026 update); available on Upstage Console, OpenRouter, BizRouter, and yupp.ai.¹¹¹²

Japanese and international variants:

• Syn Pro — Under 14B parameters; co-developed with Karakuri Inc.; launched Mar 2025; #1 on Nejumi Leaderboard for Japanese LLMs under 14B; certified by Japan METI as a domestic foundation model, Nov 2025.

Document AI and consumer products:

• Document Parse — Converts PDFs, scanned images, charts, and invoices into structured HTML and Markdown; over 95% reported OCR accuracy; available via Upstage Console API. Part of the Document Intelligence Suite (Document Parse + Information Extract) listed in the AWS Marketplace AI Agents and Tools category.
• AskUp — KakaoTalk AI chatbot combining Upstage OCR with ChatGPT; exceeded 100,000 channel friends within one week of launch; described as "ChatGPT with eyes."
• Upstage Studio — AI platform automating document-processing workflows end-to-end; launched in Japan at SushiTech Tokyo 2026.

➡️ See individual model cards for specifications, benchmarks, and API pricing.

  People

Upstage's leadership is unusually concentrated in its three co-founders, all of whom remain in executive roles. Sung Kim (CEO) is the company's public face and primary technical spokesperson; his Korean-language "Deep Learning for Everyone" lecture series made him a recognizable figure in Korea's developer community well before Upstage's founding. Hwalsuk Lee (CTO) leads model research and the engineering organization, drawing on two decades of computer vision and OCR work. Lucy Park (CSO) leads strategy and has been especially prominent in international partnership negotiations and government engagement. Benson Lee serves as Chief Business Officer, leading commercial operations, and Kasey Roh leads US business development from the San Jose office. Hiroyuki Matsushita, formerly of Panasonic and AWS, heads the Japan subsidiary.

The team draws on a talent base that spans Naver, Meta, Amazon, Nvidia, eBay, Google, Apple, and Kakao, giving it both the academic NLP pedigree (Lucy Park; MIT Technology Review Innovators Under 35 honoree) and the systems engineering depth needed for enterprise deployment at scale. Headcount is reported at under 500 people, a figure consistent with the company's emphasis on high output per researcher and its track record of achieving leaderboard-topping results with a relatively small team.

  Funding, Ownership & Business

Upstage has raised approximately KRW 400 billion (roughly $280 million reported in total) across four rounds since founding.¹ The company is private, with no audited financials publicly available; revenue and valuation figures below are reported estimates.

The Series C first tranche valued Upstage at over KRW 1 trillion (approximately $680 million estimated), granting it South Korea's first generative AI unicorn designation.¹ Analyst post-IPO valuation expectations range from KRW 2 trillion to KRW 3 trillion, though these are speculative pending a final prospectus.

Revenue is driven by four channels: API access to Solar LLMs via the Upstage Console; enterprise Document AI licensing; cloud marketplace listings on AWS and Oracle; and Korean government project contracts including the Dokpa-mo sovereign AI program. Upstage reported revenue of KRW 24.8 billion in 2025 with over 130% year-over-year growth (reported; unaudited).¹ Customers include Samsung Electronics, Korea's five largest life and property insurers, Samsung Life Insurance (OCR Pack), and multiple Korean government agencies.

  Partnerships & Ecosystem

Cloud and compute partners form the spine of Upstage's infrastructure and distribution strategy. The AWS Strategic Collaboration Agreement (June 2024) names AWS as Upstage's preferred cloud provider and lists Solar models on Amazon Bedrock, SageMaker JumpStart, and the AWS Marketplace; Amazon's participation in the August 2025 Series B Bridge deepened the commercial relationship further.⁸ Oracle Cloud Infrastructure is a training-infrastructure partner, delivering approximately 30% faster GPU cluster data transfer in ongoing collaboration. AMD became a strategic compute partner in 2025, and in March 2026 the companies announced plans to deploy MI355 GPUs for sovereign AI workloads in Korea, with AMD reported to be in talks to sell 10,000 units to Upstage (not yet confirmed as of June 2026).¹⁴ FuriosaAI partnered with Upstage in June 2025 to run Solar models on FuriosaAI's next-generation NPUs. Intel optimized Solar LLMs for Intel Core Ultra processors in June 2024.

Enterprise integrations span Korea's largest technology and consumer brands. Samsung SDS integrated Document AI and Solar LLM with its Brity Automation hyper-automation platform to improve unstructured data processing for RPA workflows. LG Electronics embedded Solar as an on-device LLM in LG Gram laptops. Polaris Office ships Solar-powered on-device AI document software. Sendbird uses Solar to power a no-code AI chatbot platform. Chosun Ilbo, Korea's largest newspaper, runs an AI-powered journalism platform on Solar (February 2025).

In Japan, Karakuri Inc. co-developed the Syn Pro LLM, and Fusic is a strategic distribution partner. Upstage launched Upstage Studio for Japanese enterprise customers at SushiTech Tokyo 2026.

  Compute & Infrastructure

Upstage does not operate its own data centers at hyperscaler scale; instead it has built a multi-cloud compute strategy anchored to AWS for training and inference, with Oracle Cloud Infrastructure providing supplementary high-speed GPU cluster connectivity. The AWS relationship was formalized as a Strategic Collaboration Agreement in June 2024 and reinforced by Amazon's direct investment in the August 2025 Series B Bridge, giving Upstage preferential access to capacity and tighter integration with Amazon Bedrock.⁸

For sovereign AI infrastructure in Korea, the company is shifting toward AMD MI355 GPUs, with talks underway as of March 2026 to procure up to 10,000 units — a significant expansion of domestic compute independence from US hyperscaler infrastructure.¹⁴ The FuriosaAI partnership (June 2025) opens an additional vector toward inference on Korean-designed AI chips. Training scale for Solar Open 100B was disclosed as 19.7 trillion tokens of pre-training, placing it in the same order of magnitude as contemporary frontier pre-training runs while relying on cloud-burst compute rather than owned hardware.

  Notable Events & Controversies

Solar Open 100B plagiarism allegations (January 2026). Shortly after the December 2025 release of Solar Open 100B, Saionic AI CEO Ko Seok-hyun published a GitHub analysis reporting that the model exhibited 96.8% similarity to the Chinese model GLM 4.5 Air at the LayerNorm layer — effectively alleging that Upstage's flagship open model was a copy of a foreign model rather than a domestically built system.¹⁵ The allegations carried particular weight because Upstage's sovereign AI credentials underpin both its participation in the Korean government's Dokpa-mo project (worth KRW 213.6 billion in grants) and the credibility of its planned KOSPI IPO. Upstage CEO Sung Kim responded by holding a public online verification session, disclosing training logs and checkpoints, and presenting a reanalysis using Pearson correlation coefficients that showed no structural match to GLM 4.5 Air.¹⁵ The Korean AI research community remained divided, but Upstage passed the first-phase Dokpa-mo evaluation in January 2026 alongside LG AI Research and SK Telecom, and its unicorn fundraise in April 2026 proceeded on schedule, suggesting that investors accepted the company's rebuttal.

  Competitive Position

Upstage holds an estimated 35% share of South Korea's private LLM market (reported; no independent audit available). Its primary Korean competitors are Naver HyperCLOVA X, KT AI, LG AI Research EXAONE, and SK Telecom A. Notably, Naver and NC AI were eliminated in the first phase of the Dokpa-mo sovereign AI project evaluation, which Upstage passed — a competitive signal in the context of Korea's government AI procurement.¹⁵ Globally, Upstage competes in the enterprise segment against OpenAI, Anthropic, Mistral, Qwen (Alibaba), and DeepSeek, though it deliberately avoids head-to-head competition with hyperscaler flagship models and instead targets the enterprise deployment layer where document-processing accuracy, cost efficiency, and data sovereignty matter most.

Upstage's four principal points of differentiation are: (1) the DUS and SnapPO methodologies, enabling frontier-class performance at dramatically reduced active parameter counts; (2) a deep document intelligence stack tuned for regulated-industry workflows where raw language model performance is insufficient without accurate OCR and structured extraction; (3) Korean and Japanese language expertise that domestic and regional enterprises cannot match with generic English-first models; and (4) sovereign AI positioning that aligns with Korean government policy and appeals to APAC enterprises wary of data sovereignty risks with US providers.

  Outlook & Roadmap

Upstage's near-term priorities converge on four tracks. First, the Daum acquisition (formal agreement signed 7 May 2026) would transform Upstage from a model API provider into a platform company integrating search, content, and agentic AI — using Solar LLM to power Daum's portal and leveraging Daum's content corpus as training and retrieval data.¹³ The all-stock deal's final terms and the implied Daum valuation have not been disclosed.

Second, the Dokpa-mo Sovereign AI Foundation Model Project: Upstage is competing in the second phase (through June 2026) alongside LG AI Research and SK Telecom, with a final evaluation scheduled for December 2026 that will determine allocation of KRW 213.6 billion in government AI infrastructure funding. Outcome is not yet known.

Third, the planned KOSPI IPO, with KB Securities and Mirae Asset as lead underwriters, is targeted for the second half of 2026 or first half of 2027. Analyst estimates project a post-IPO market capitalization of KRW 2–3 trillion, though these figures are speculative and subject to market conditions and the IPO prospectus review process.⁹

Fourth, model development continues with Solar Pro 3 in active deployment and at least one additional planned release: a Solar Pro 1.5 model reported in one source as a 33B multimodal system (mentioned in a single source as of April 2026 and not confirmed by the official Upstage newsroom).³ Further Syn Japanese-language variants are also expected following METI certification of Syn Pro.

Strategically, Upstage is positioning itself for a transition from Korea's leading LLM API provider to a full AI platform company with integrated search, content, and document intelligence capabilities — a trajectory that, if the Daum acquisition closes and the IPO proceeds, would make it the most significant domestically built AI platform in South Korea.

---

  References

1. Upstage becomes first Korean AI service startup to reach unicorn status — Seoul Economic Daily
2. Upstage founding story — Upstage Blog
3. Upstage — Wikipedia
4. SOLAR 10.7B: Scaling Large Language Models with Simple yet Effective Depth Up-Scaling — arXiv:2312.15166
5. Upstage Series B Funding — Upstage Newsroom
6. Upstage Solar Pro — Seoulz
7. Solar Pro 2 announcement — Upstage Newsroom
8. Upstage bags $45M to expand AI-native document intelligence suite — SiliconAngle
9. AI startup Upstage hires KB Securities, Mirae for Korea IPO — Bloomberg
10. Solar Open 100B — Hugging Face Model Card
11. Solar Pro 3 release blog — Upstage
12. Upstage Solar Pro 3 agentic update — Upstage (March 2026)
13. Upstage agrees to acquire Daum from Kakao — Korea Economic Daily Global
14. Upstage deepens partnership with AMD to power sovereign AI push — Korea Times
15. Upstage refutes Chinese AI model plagiarism claims — Seoul Economic Daily
16. Upstage Series B Bridge — Upstage Newsroom

  References

1. Upstage unicorn announcement, Series C first tranche, April 2026 — Seoul Economic Daily. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. Upstage founding story and team background — Upstage Blog. ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Background details, leaderboard milestones, Solar Pro 1.5 mention — Wikipedia: Upstage; Seoulz. ↩ ↩² ↩³

4. SOLAR 10.7B paper — arXiv:2312.15166. ↩ ↩² ↩³

5. Series B funding — Upstage Newsroom. ↩

6. Solar Pro GA on AWS — Seoulz. ↩ ↩²

7. Solar Pro 2 release — Upstage Newsroom. ↩ ↩²

8. Series B Bridge and AWS/AMD investment — SiliconAngle; Upstage Newsroom. ↩ ↩² ↩³

9. IPO underwriter announcement — Bloomberg. ↩ ↩²

10. Solar Open 100B model card — Hugging Face. ↩ ↩²

11. Solar Pro 3 release — Upstage Blog. ↩ ↩² ↩³

12. Solar Pro 3 March 2026 agentic update — Upstage Blog. ↩ ↩²

13. Daum acquisition announcement — Korea Economic Daily Global. ↩ ↩²

14. AMD MI355 partnership — Korea Times. ↩ ↩² ↩³

15. Plagiarism controversy and Dokpa-mo evaluation — Seoul Economic Daily. ↩ ↩² ↩³