← Analysis archivereport.jsonevidence.jsonDossier →
Upstage (Solar)Neolabgenerated Jun 27, 2026 · 6h

Upstage (Solar) analysis

36 cited sources241 signals on file

Thesis

Upstage is executing a platform consolidation play, moving from a pure models-and-APIs company toward an integrated AI stack combining proprietary Solar LLMs, acquired portal (Daum) assets, and agentic platforms (Timely).

Core thesis: Upstage is building an AI-for-everyone ecosystem — not just enterprise AI, but consumer-facing AI through a portal. The Solar model family (10.7B→31B→Open 100B→22B in preview/pro) shows the company can match frontier-scale performance at small parameters, and the company is pushing into sovereign AI infrastructure with AMD, Furiosa, and national foundation model projects [W1, W2, W3, W5, W6].

Signal Desks

Hiring

No cited evidence in this pack for hiring. The evidence pack contains no direct job postings or role listings. All hiring signals come from the company's own news posts and blog content about team expansion, location strategy, and commercialization buildout.

Gap analysis: There is no cited hiring evidence. The gap in direct hiring signals suggests the company has not published open roles. However, the company's blog and news posts describe its ambassador program, bootcamp projects, hackathon events, and SW maestro scaffold submissions as community-building activities [E14, E18, E19, E22, E27].

Forks

  • Hermes-agent (NousResearch): Agent framework for function-calling agents — early signal of agentic workflow building, preceding a potential release or hiring wave E15.
  • Open-webui (open-webui): Deployment of chatbot/UI infrastructure — suggests the company is testing chat interfaces and deployment tooling E44.
  • vLLM (vllm-project): Inference serving engine — core infrastructure for model deployment at scale E45.
  • GPU scheduling (AliyunContainerService): GPU sharing and scheduling infrastructure for Kubernetes — suggests the company is optimizing compute resource allocation E46.
  • Code/security (anomalyco): Code analysis and security auditing — suggests the company is inspecting upstream tooling for vulnerabilities E50.
  • Training environments (NVIDIA-NeMo): RL training and simulation environments — suggests the company is building on NVIDIA's simulation and training infrastructure E51.
  • Core ML (huggingface/transformers): Core ML dependency and model architecture — the company directly depends on Hugging Face transformers for model training and fine-tuning E53.
  • Agent evaluation (sierra-research/tau2-bench): Agent benchmark evaluation — the company is testing and benchmarking agentic capabilities against frontier models E57.
  • Observability (open-telemetry): Telemetry and observability collection — suggests the company is building out monitoring and observability infrastructure E59.
  • Inference optimization (coreweave/tensorizer): Tensorizer and inference optimization for serving — suggests the company is optimizing model serving and inference at the edge E60.

Releases

Key model releases (Hugging Face):

  • SOLAR-10.7B-Instruct-v1.0 (Dec 2023): First instruct model release. 10.7B parameters, 63.5k downloads, 656 likes E1.
  • Solar-Open-100B (Dec 2025): Open-source 100B release. 102.6B parameters, 5k downloads, 479 likes E2.
  • solar-pro-preview-instruct (Sep 2024): Preview instruct release. 22B parameters, 42.7k downloads, 457 likes E3.
  • SOLAR-10.7B-v1.0 (Dec 2023): Base model release. 10.7B parameters, 12k downloads, 321 likes E4.
  • SOLAR-0-70b-16bit (Jul 2023): 70B-16bit release. 13k downloads, 259 likes E5.
  • TinySolar-187m-4k (May 2026): Tiny model release. 187M parameters, 76 downloads E16.
  • TinySolar-111m-4k (May 2026): Tiny model release. 111M parameters, 20 downloads E17.
  • llama-30b-instruct (Jul 2023): LLaMA instruct fine-tune. 900 downloads E21.
  • llama-65b-instruct (Jul 2023): LLaMA instruct fine-tune. 805 downloads E23.
  • solar-1-mini-tokenizer (May 2024): Tokenizer release E24.
  • TinySolar-248m-4k (Feb 2024): Tiny base model. 417 downloads E25.
  • solar-pro-preview-tokenizer (Sep 2024): Tokenizer release E26.
  • TinySolar-248m-4k-code-instruct (Apr 2024): Code instruct fine-tune. 107 downloads E28.
  • solar-docvision-preview-tokenizer (Sep 2024): DocVision tokenizer release E29.
  • TinySolar-248m-4k-py (Feb 2024): Python base model. 106 downloads E30.
  • solar-pro-tokenizer (Nov 2024): Tokenizer release E31.
  • TinySolar-248m-4k-py-instruct (Apr 2024): Python instruct fine-tune. 68 downloads E33.
  • TFLOP (Nov 2025): Table structure recognition framework. 24 downloads E34.
  • solar-pro3-tokenizer (Jan 2026): Pro3 tokenizer E35.
  • solar-pro2-tokenizer (Jul 2025): Pro2 tokenizer E39.

Other releases (tokens, TinySolar base models): TinySolar-248m-4k, TinySolar-248m-4k-code-instruct, TinySolar-248m-4k-py, TinySolar-248m-4k-py-instruct [E25, E28, E30, E33].

Platform products:

  • Evalverse (GitHub): LLM evaluation framework. 235 stars E37.
  • Cookbook (GitHub): API examples and guides. 199 stars E40.
  • TFLOP (GitHub): Official implementation of table structure recognition. 51 stars E41.
  • Dataverse (GitHub): Data science and engineering universe. 563 stars E32.
  • Security-audit (GitHub): Public security auditing. E38.
  • Solar-prompt-cookbook (GitHub): Prompt engineering cookbook. 252 stars E36.

Talking

Blog posts and news announcements:

  • Modular AI Tech Stack (P1): The company is explaining how to build for a world where models change every year — modular stack for insurance. GTM: enterprise insurance workflow.
  • AskUp Learn English (P2): Consumer use case for learning English via AskUp chatbot powered by Solar (LLaMA-2 fine-tune). Consumer app.
  • OCR API Free Trial Event (P3): Launch event for the OCR API. Cloud-based OCR for insurance claims, resume review, receipt expense processing, and image translation.
  • Document Parse Enhanced (P4): VLM-based document parsing for complex tables, checkboxes, charts, and diagrams. Document AI.
  • Solar Pro 2 (P5, P21): Next-gen LLM with 31B parameters, hybrid Chat/Reasoning modes, CoT reasoning, MMLU-Pro, Math500, AIME, SWE-Bench benchmarks. Outperforms GPT-4o, DeepSeek R1, Mistral Small 3.2, Alibaba Qwen 3. Korean language understanding, Arena-Hard-Auto, Hae-Rae, Ko-MMLU. Finance, medicine, law domains. Agentic AI executing multi-step tasks with external tools.
  • Startup Branding (P6): Why startups need branding — consistency, brand identity, and unique color.
  • Underwriting Reinvention (P7): 90-day plan phase 1 — diagnose and elevate. Insurance carriers and MGAs. AI Space platform. Submission workflow mapping. Automation vs. human expertise.
  • Buy to Build (P8): Modernization playbook. Agentic Information Extract. Document intelligence. Legacy systems as liabilities.
  • When Ontology Moves Faster Than IT (P11): Schema as configuration, not trained model. Few-shot terminology examples. Confidence scores. Source tracing. Auto-improvement feedback loops.
  • AI Ethics (P12): AI ethics and corporate efforts. AI safety, human rights, freedom, responsibility, creativity, social inequality. AI Act.
  • AI Talk Show Recap (P13): Community ecosystem for LLMs. AI Talk Show in Seoul with SBVA. Richard Socher (you.com), Cindy Jin (SBVA), Sung Kim (Upstage CEO).
  • OCR Information Extraction (P14): Digital assetization. Document OCR. Detector/recognizer. Deep learning models.
  • AskUp User Guide (P15): KakaoTalk channel. 100k users within a week. 300k registered users. Generative AI mascot. Upsketch image generation. AskUp Biz (B2B). AskUp Doc, AskUp Web, AskUp Slack.
  • Evalverse (P16): LLM evaluation. Unified framework, lm-evaluation-harness, FastChat. No-code Slack bot. Hugging Face hub.
  • Solar API Beta (P17): Free API beta. Solar LLM Innovators Award. $200k credits. English-Korean translation. GPT-4, DeepL context-aware.
  • AWS SCA Global (P18): Strategic collaboration with AWS. Co-sales, GTM. Minority investment. SageMaker, Trainium, Inferentia. Public sector AI.
  • LLM Evaluation Part 1 (P25): Benchmark datasets. Perplexity, BLEU. SAT for LLMs.
  • Solar LLM with Predibase (P26): Fine-tuning beats GPT-4. Predibase fine-tuning leaderboard. 500 experiments.
  • ICDAR Win Interview (P27): ICDAR 2023 competition. HierText, VQAonBD, IHTR. Multimodal AI researcher.
  • 2023 AI Tech Trend Seargest (P20): Search + suggest. Personalized search and recommendation. YouTube, OTT, Naver. AI shopping curation.
  • Solar Pro 2 Frontier (P21): Artificial Analysis Intelligence Index. Top 10 Frontier Models. Grok-4, Kimi-K2, DeepSeek V3.
  • Solar Sendbird (P22): No-code AI chatbot. 300M users, 4,000 enterprise companies. Customer support and engagement.
  • Kbank MOU (P23): LLM for finance. KT, kt cloud, K Bank. GPU infrastructure. AI hallucinations.
  • Solar LLM for Writing Project Proposals (P24): Tips for effective project proposals.
  • LLM Evaluation Part 1 (P25): Benchmark datasets.
  • Solar LLM with Predibase (P26): Fine-tuning.
  • ICDAR Win Interview (P27): Competition results.

Category Implications

Connecting cited evidence to category-specific strategy, infrastructure, product, research, hiring, and GTM implications.

Traction Highlights

No direct traction evidence (stars, forks) in this pack for traction/community growth. The company's GitHub stars and Hugging Face likes provide community validation, but no direct user metrics are available.