ORNITH 1.0 35B : Meilleur que Qwen 3.6 35B ?

ORNITH 1.0 35B : Meilleur que Qwen 3.6 35B ?

ORNITH 1.0 35B: Better than Qwen 3.6 35B?

Title history
  • ORNITH 1.0 35B : Meilleur que Qwen 3.6 35B ? — observed on 20261101-0107 UTC (YouTube)

by Nathan Inline · View the channel on PSI

Applied Sciences & Engineering AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Nathan Inline 👥 4K 📅 July 9, 2026 ⏱ 18 min 👁 3K 📄 opinion experte 🧭 2026-11-01
Available in: English (current) Français

Keywords

benchmarkcodinglocal modelagentself-improvement

Summary

Nathan Inline tests the newly released Ornith 1.0 35B model, comparing it against Qwen 3.6 35B and other models. He first reviews benchmark results that suggest Ornith outperforms many competitors. He then demonstrates practical usage through his own app Jean, including memory functions and web access. The model successfully retrieves subscriber counts and explains the novel ’learn to learn’ training approach, which shapes its own problem-solving strategies. However, in a coding test (creating a basketball game), the model fails to produce a functional result and also fails to correct itself after multiple attempts. The creator concludes that while Ornith 1.0 35B excels at web research and agentic tasks, it is inferior to Qwen 3.6 27B for coding. The video also heavily promotes the creator’s products, Jean and Agent Link, with a promotional segment of approximately 2 minutes. Overall, the video provides a practical but subjective evaluation of a new AI model.

151 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides a clear, step-by-step evaluation of Ornith 1.0 35B, combining benchmark analysis with hands-on tests. The argumentation is straightforward: benchmarks are taken at face value but then checked through real-world tasks. The memory and web-browsing demonstrations effectively show the model’s strengths for agentic use. The coding test reveals significant limitations, providing a balanced view. However, the creator’s conclusions rely on his own tests, which lack controlled conditions or reproducibility details. The reasoning is logical, but the absence of quantitative metrics (e.g., success rates, errors) weakens the scientific strength.

Scientific Rigor, Source Quality, Title Accuracy

The sources cited are the Hugging Face model page, which hosts the official benchmarks and details, and the creator’s own applications (Jean and Agent Link). No independent scholarly sources are referenced. The title’s claim about Qwen 3.6 35B is directly addressed in the conclusion, so title and content match. A promotional segment for Agent Link is present, but it does not interfere with the core testing. No user comments were provided for analysis.

178 words

Title / Content Match

The title accurately reflects the content: the video tests whether Ornith 1.0 35B outperforms Qwen 3.6 35B based on benchmarks and practical tasks.

Quality & Reliability

6/10

The video offers a practical, hands-on test of a new AI model, but the evaluation is subjective and not methodologically rigorous. The creator promotes his own tools, which reduces overall neutrality. Benchmark comparisons are based on claimed results rather than independent verification.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

Contribution & Novelties

This video contributes a practical, real-world evaluation of Ornith 1.0 35B, bridging benchmark scores and actual usability. It highlights the model’s unique ’learn to learn’ training paradigm, which distinguishes it from fixed-scaffold models like Claude. The creator’s hands-on testing reveals that while the model excels at agentic tasks, it falls short in code generation and debugging, providing nuance to the impressive benchmarks. The promotion of his own tools adds a practical angle for users interested in local AI deployment.

Pour aller plus loin :

  • Meta-learning (computer science) — Explores the concept of learning to learn, directly relevant to Ornith’s training approach.
  • Reinforcement learning — The backbone of reward-based training used to optimize scaffolds in Ornith.
  • Agentic AI — Discusses AI systems that autonomously plan and execute tasks, as Ornith is designed to do.

133 words

Radar Profile

The radar profile shows strong information quantity but moderate quality and reliability, reflecting the video's informal testing methodology. Technical level is high, appealing to intermediate users, but the overall trust is tempered by subjective opinions and promotional content.

Reliability 6/10