Solar-Open-100B vs GLM-4.5-Air
Model Derivation Forensics

Controversy resolved through proactive verification by Upstage and the community

• • Reading time: ~20 min • 한국어
✅ Controversy Resolved: Accuser apologized and repository closed

Executive Summary

In December 2025, Sionic AI raised allegations that Upstage's Solar-Open-100B model was derived from Zhipu AI's GLM-4.5-Air, sparking a major controversy in the Korean AI industry. The key evidence was a statistical analysis showing that the Cosine similarity of LayerNorm weights between the two models reached 0.989 (182 sigma).

In response, Professor Hyunwoong Ko demonstrated that similar similarity levels also appeared with a third model (Microsoft Phi-3.5-MoE), arguing this could be explained by structural convergence in RMSNorm. Subsequently, Professor Kyunghyun Cho (NYU) conducted a Pearson correlation analysis revealing the actual correlation between the two models was only 0.0054, delivering a verdict of "Weak Evidence, Inconclusive."

On January 2, 2026, Upstage's public verification session revealed W&B training logs and checkpoints, proving independent development at the 99.9996% level. Afterward, Sionic AI's CEO posted an official apology and closed the original repository, completely resolving the controversy. This case stands as an exemplary demonstration of the importance of scientific methodology in model forensics and transparent verification in the era of LLM open weights.

Summary (Executive Summary)

This report examines the controversy over whether South Korean AI company Upstage's Solar-Open-100B model was derived from Chinese Zhipu AI's GLM-4.5-Air model. The dispute progressed from initial allegations, through circumstantial rebuttals, to a technical resolution via Professor Kyunghyun Cho (NYU)'s statistical verification.

Phase Actor Core Claim Situation
1. Allegation Sionic AI "The code is identical, and the weight similarity is 0.99. This is 100% copied." 🚨 Solar in crisis
2. Rebuttal Hyunwoong Ko "MS Phi-2 has the same code too. It's just a standard template." 🛡️ Counterattack begins
3. Verification Prof. Kyunghyun Cho "The 0.99 weight similarity is merely a mathematical artifact. The actual pattern (Pearson) is different." ⚖️ Resolved

Final Conclusion: Professor Kyunghyun Cho's Pearson correlation analysis yielded 0.0054 — statistically proving that the weight patterns of the two models have no correlation whatsoever (independent training). The accuser subsequently posted an apology and closed the repository.

"Weak Evidence, Inconclusive" — Sionic AI's claims are insufficient and inconclusive.

— Professor Kyunghyun Cho's final verdict

Background

Models Under Analysis

The key players in this controversy are Solar-Open-100B (Upstage) and GLM-4.5-Air (Zhipu AI), along with the control model Phi-3.5-MoE (Microsoft). All three models adopt the MoE architecture and RMSNorm, making this structural commonality a key variable in the debate.

Model Name Developer Parameters Architecture Features
Solar-Open-100B Upstage (South Korea) ~100B MoE, 48 Layers, Presumed DUS technique
GLM-4.5-Air Zhipu AI (China) 106B MoE, 46 Layers, 96 Heads
Phi-3.5-MoE Microsoft (USA) ~42B MoE, 32 Layers, Control model

Sources of the Controversy

Derivation Hypothesis (Sionic AI)

sionic-ai/solar-vs-glm

Focused on statistical outlier detection

Independence Hypothesis (Hyunwoongko)

hyunwoongko/solar-vs-glm-vs-phi

Focused on metric validity verification

Claim Reconstruction and Comparison

The table below compares the core claims, key evidence, and blind spots of the derivation hypothesis (Sionic AI) and the independence hypothesis (Hyunwoong Ko). Sionic AI centers on the "182-sigma outlier" in LayerNorm, while Professor Hyunwoong Ko argues for "RMSNorm structural convergence."

Category Sionic AI (Derivation Hypothesis) Hyunwoongko (Independence Hypothesis)
Core Claim Solar is derived from GLM, evidenced by abnormal same-layer LayerNorm similarity (≈0.989) and selective preservation RMSNorm has low information content and strong structural constraints, so high Cosine similarity is an architectural inevitability (convergence), not evidence of plagiarism
Key Evidence • Same-layer LN Cosine ≈ 0.99
• 182σ above internal model baseline (0.377)
• "Selective preservation" pattern 		• 0.9+ similarity also occurs between Phi-3.5-MoE and GLM
• LN similarity of 0.99 observed across other models
• RMSNorm vector direction bias characteristics
Blind Spots Decisive evidence relies excessively on low-information parameters (LayerNorm). Mathematical properties of RMSNorm not considered Does not directly refute the near-zero Attention/MoE similarity (evidence of retraining). No explanation for why "selective preservation" occurs

Key Questions Requiring Cross-Verification

The key to resolving this controversy is whether similarity remains high in "high-information tensors (Attention/MoE)" as well. The derivation hypothesis asks "Does the selective preservation pattern hold?", while the independence hypothesis asks "Is Solar-GLM significantly closer than Solar-Phi?" The two questions below essentially require the same experiment, and the results will either strengthen the derivation hypothesis or confirm the independence hypothesis.

Derivation Verification Question

"Does the selective preservation pattern (LN is identical, everything else differs) persist when we remove RMSNorm's mathematical properties and compare high-information tensors (Attention Weights)?"

Independence Verification Question

"When excluding RMSNorm, is the distance between Solar and GLM statistically significantly closer than the distance between Solar and Phi?"

Combined Verification Framework: Retain Sionic's "selective preservation" hypothesis, but change the target variables from RMSNorm to Attention / MoE / FFN and perform a three-way comparison of Solar-GLM-Phi.

Outcome A (Derivation Strengthened)

Solar-GLM similarity in high-information tensors is significantly higher than with other models

Outcome B (Independence Confirmed)

Similarity across all models in high-information tensors is low or comparable

Detailed Data Analysis of Each Claim

Derivation Hypothesis: "The Shock of 182 Sigma"

Sionic AI's analysis reveals extreme similarity differences across tensor types. The table below shows the per-tensor Cosine similarity between Solar-Open-100B and GLM-4.5-Air.

Tensor Type Cosine Similarity Sionic's Interpretation Notes
input_layernorm 0.949 Original preserved Structural skeleton
post_attention_layernorm 0.986 Original preserved Structural skeleton
k_proj / v_proj 0.001 Retrained Attention (intelligence)
mlp.gate 0.004 Retrained MoE Router (intelligence)
embed_tokens 0.002 Retrained Vocabulary expansion

Sionic AI's Logic

If the models were independently trained, a 0.99 similarity could not occur by chance (P-value < 10^(-1000)). Therefore, they took the skeleton (LN) and swapped out only the muscles (Attn/MLP).

Independence Hypothesis: "The RMSNorm Trap"

Professor Hyunwoong Ko introduced a third model, Phi-3.5-MoE, to prove that this "similarity" is an illusion. Below is the LayerNorm Cosine similarity matrix for the 10th layer.

Model Solar-Open GLM-4.5-Air Phi-3.5-MoE
Solar-Open 1.000 0.9+ 0.9+
GLM-4.5-Air 0.9+ 1.000 0.9+
Phi-3.5-MoE 0.9+ 0.9+ 1.000

Key Observation

Even between completely different models (Phi vs GLM), LayerNorm similarity exceeds 0.9.

Cause Analysis

As training progresses, RMSNorm converges toward specific scale values, causing vector directions to become similar -- a phenomenon known as "structural convergence." Therefore, LN similarity cannot serve as a model's "fingerprint."

Professor Kyunghyun Cho's Final Verdict

World-renowned AI scholar Professor Kyunghyun Cho (NYU) reanalyzed the central issue of "cosine similarity 0.99." The comparison cards below show the difference between Raw Cosine (0.99, illusion) and Pearson Correlation (0.0054, actual pattern), while the final verdict table delivers "Not guilty (industry practice)" and "Not guilty (statistically proven)" for the code and weight issues respectively.

Detailed Data Analysis

Raw Cosine (The Truth Behind 0.99)

  • LayerNorm/RMSNorm weights inherently have values close to 1.0
  • Even randomly comparing different layers yields a similarity of 0.95
  • In other words, 0.99 is a "naturally high value", not evidence of copying (Uninformative)

Pearson Correlation (Evidence of 0.00)

  • Measures how similar the weight "patterns" are, not absolute magnitude
  • Result: 0.0054
  • Indicates the weight patterns of the two models have no correlation at all (independent)

📊 What is Pearson Correlation? (Beginner's Guide)

The Pearson correlation coefficient measures whether two datasets "move in the same pattern." When the value is close to 1, they "rise and fall together"; when close to 0, they "move independently of each other."

Example: High correlation (r ≈ 0.99)

Alice and Bob's test scores in 5 subjects:

→ The two lines move in nearly identical patterns
→ "Same pattern" = Possible copying!

Example: No correlation (r ≈ 0)

Charlie and Diana's test scores in 5 subjects:

→ The two lines move independently
→ "Different patterns" = Each studied on their own!

Solar vs GLM Analysis Result: Pearson correlation of the two models' weights = 0.0054 (nearly 0) → The two models were "trained with different patterns" = independently developed

0.99
Raw Cosine
(Illusion)
0.0054
Pearson
(Actual Pattern)
0%
Correlation
(Independent)
✓
Not Guilty
(Statistically Proven)

"Weak Evidence, Inconclusive"

Sionic AI's claims are insufficient evidence / inconclusive.

"Don't be fooled by the number 0.99. When examined statistically, the weights of the two models appear to have been trained independently of each other."

Issue Accusation (Copied!) Defense & Verification (Not true!) Final Verdict
Code "The code is identical to GLM, down to every detail!" "MS Phi-2 is identical too. It's because they used an industry-standard template." ✅ Not Guilty (Industry Practice)
Weights "The vector cosine similarity is 0.99! Case closed!" [Prof. Kyunghyun Cho] "LayerNorm inherently produces high values. The pattern (Pearson) shows 0.00." ✅ Not Guilty (Statistically Proven)

Conclusion: The Essence and Resolution of the Controversy

This controversy can be summarized as an incident stemming from "misunderstanding of open-source template usage" and "insufficient understanding of RMSNorm's mathematical properties."

Three-Line Summary:

  1. 1. Sionic AI: "Numbers (0.99) don't lie! They copied it!"
  2. 2. Hyunwoong Ko: "MS's code is identical too. Stop making false accusations."
  3. 3. Prof. Kyunghyun Cho: "0.99 is an artifact. When examined statistically, they are indeed different models. Case closed."

Technical Conclusion

Professor Kyunghyun Cho's Pearson correlation analysis statistically rejected Sionic AI's claims. Solar-Open-100B is a model trained independently from GLM-4.5-Air.

Epilogue: Resolution of the Controversy

Beyond technical analysis, this controversy was completely resolved through concrete actions. The timeline below documents three decisive events: Upstage's public verification (1/2), Sionic AI CEO's official apology (1/3), and the closure of the original repository (currently 404).

🎤

Upstage's Public Verification

2026.01.02
  • Held a public verification session in Gangnam, Seoul
  • CEO Sunghun Kim + core engineers + external experts attended
  • Disclosed W&B training logs and intermediate checkpoints
  • Only 0.0004% of total parameters showed similarity → Proved 99.9996% independent development
🙏

Accuser's Apology

2026.01.03
  • Sionic AI CEO Seokhyun Ko posted an apology via LinkedIn
  • "I raised allegations without sufficient verification"
  • "LayerNorm cosine similarity alone is insufficient to determine model weight reuse"
  • Official apology to Upstage CEO Sunghun Kim and team members
🗂️

Original Repository Closed

Current

The sionic-ai/solar-vs-glm repository currently returns 404 Not Found. The analysis materials that served as the basis for the allegations have been withdrawn.

Controversy Resolution Summary

✅
Technically Not Guilty
Prof. Cho's verification
🎤
Public Proof
Upstage verification session
🙏
Official Apology
Accuser's acknowledgment

Lessons from This Controversy

The "Solar vs GLM" incident demonstrated the importance of "Model Forensics" in the era of LLM Open Weights, as well as serving as an exemplary case of healthy dispute resolution culture.

Lessons for Analysts

Hasty conclusions based on a single metric (especially low-information parameters like LayerNorm) are dangerous. Public accusations without sufficient verification carry great responsibility.

Lessons for Developers

Transparency in model development processes (WandB, Checkpoints) is no longer optional but "insurance." Upstage's proactive public verification stands as an exemplary response.

Lessons for the Community

Asking "which experiment is scientifically valid" rather than "who is right" creates a healthier ecosystem. Resolving through public verification and dialogue rather than legal conflict demonstrates the maturity of the AI ecosystem.

The true winner of this controversy was not the side with the loudest voice, but "the side that transparently proved their case through scientific verification." The accuser's willingness to acknowledge their mistake and apologize also deserves respect.

Frequently Asked Questions (FAQ)

What is the '182 sigma' that Sionic AI claims?
Sionic AI claims that the Cosine similarity between same-layer LayerNorm of Solar-Open-100B and GLM-4.5-Air is approximately 0.989, which is more than 182 standard deviations above the internal model baseline (0.377). This means the probability of this occurring by chance is less than 10^(-1000).
What is the core of Professor Hyunwoong Ko's rebuttal?
Professor Hyunwoong Ko demonstrated that 0.9+ similarity also occurs between a third model, Phi-3.5-MoE, and GLM. Since RMSNorm exhibits "structural convergence" toward specific values as training progresses, high LN similarity is an architectural inevitability, not evidence of plagiarism.
What is the selective preservation pattern?
A pattern discovered by Sionic AI where LayerNorm weights show high similarity at the 0.99 level, while Attention (k_proj, v_proj) and MoE (mlp.gate) weights show very low similarity at 0.001-0.004 levels. This was the basis for the interpretation that "the skeleton (LN) was taken and only the muscles (Attn/MLP) were replaced."
Why can't LayerNorm alone be used to make a judgment?
RMSNorm (a variant of LayerNorm) has low information content and strong structural constraints. It has the property of converging toward specific scale values as training progresses, which can result in high similarity even between different models. Therefore, comparison of high-information tensors (Attention, MoE) that carry the model's "intelligence" is necessary.
How can a final verdict be reached on this controversy?
The key question is "Does the selective preservation pattern also reproduce in high-information tensors (Attention/MoE)?" If Solar-GLM similarity in high-information tensors is significantly higher than with other models, the derivation hypothesis is strengthened. If similarity across all models is low or comparable, the independence hypothesis is confirmed.
How was the controversy resolved?
On January 2, 2026, Upstage held a public verification session in Gangnam, disclosing W&B training logs and checkpoints. After proving 99.9996% independent development, Sionic AI CEO Seokhyun Ko posted an official apology on LinkedIn on January 3 and closed the original repository (sionic-ai/solar-vs-glm).
What lessons does this controversy offer the AI industry?
It demonstrates the importance of "model forensics" in the era of LLM open weights. Analysts should be wary of hasty conclusions based on a single metric (especially low-information parameters), and developers should maintain transparency in their training processes (WandB, Checkpoints) as insurance. This case has also become an exemplary model of a mature AI ecosystem that resolves disputes through public verification and dialogue rather than legal conflict.

Report Download

📑

Forensic Analysis of the Solar-Open-100B vs GLM-4.5-Air Model Derivation Controversy

182-sigma LayerNorm Similarity vs Structural Convergence -- A Scientific Verification Framework

Pebblous Data Communication Team | January 2026

View Original (v2) Download

References

  1. [1] Solar-Open-100B vs GLM-4.5-Air: Weight Derivation Analysis (Derivation Hypothesis) - Sionic AI GitHub. [404 - Repository Closed]
  2. [2] Solar vs GLM vs Phi: Structural Convergence Rebuttal (Independence Hypothesis) - Hyunwoong Ko GitHub. https://github.com/hyunwoongko/solar-vs-glm-vs-phi
  3. [3] Upstage Controversy Ends with Apology from Sionic AI CEO - KoreaTechDesk. https://koreatechdesk.com/...
  4. [4] Upstage Pushes Back Against Claims of Copying Chinese AI Model - KMJ. https://www.kmjournal.net/...
  5. [5] South Korea's Upstage Faces Plagiarism Claims Over Solar Open Model. Public Data Review Tells a Different Story - KMJ. https://www.kmjournal.net/...
  6. [6] AI Independence on Trial: How the Upstage Defense Tests Korea's Sovereign Model Strategy - KoreaTechDesk. https://koreatechdesk.com/...
  7. [7] Upstage Denies Copying Chinese Model, Vows Public Verification - Chosun. https://www.chosun.com/english/...
  8. [8] Upstage Solar-Open-100B Public Validation - LocalLLaMA - Reddit. https://www.reddit.com/r/LocalLLaMA/...

Disclaimer: This report is based on data from two publicly available GitHub repositories (sionic-ai, hyunwoongko) and does not conclusively determine the veracity of any particular claim. The purpose of this report is to propose a methodology for scientific verification.