Scaling AI Isn’t Everything. Nvidia and a Hong Kong University Unveil a Model Stronger Than Its Rivals.

Scaling AI Isn’t Everything. Nvidia and a Hong Kong University Unveil a Model Stronger Than Its Rivals.

Ondřej Barták
Ondřej Barták
Entrepreneur and Programmer
2. 12. 2025
1 minutes reading · 7 views
Scaling AI Isn’t Everything. Nvidia and a Hong Kong University Unveil a Model Stronger Than Its Rivals.

Imagine a small model with 8 billion parameters called Orchestrator, which acts like a conductor in an orchestra full of tools and intelligent models. This creation from researchers at Nvidia and the University of Hong Kong solves complex tasks, such as those from Humanity's Last Exam (HLE), where it achieved a score of 37.1%, while GPT-5 scored only 35.1%. And all that at 2.5 times lower cost! No giant model, just smart coordination.

Why are small models more powerful than large ones?

Orchestrator is no lone hero—it calls on tools such as the Tavily search API for web searches, a Python sandbox for running code, or specialized models such as Qwen2.5-Math-72B for mathematics. During training, it uses reinforcement learning with rewards for correct results, low costs, and adherence to user preferences. For example, on the FRAMES benchmark, it outperformed GPT-5 with a success rate of 76.3% at just 30% of the cost.

Benchmark results and cost
Benchmark results and cost

In each round, Orchestrator reasons, selects a tool—such as GPT-5-mini for coding or Llama-3.3-70B-Instruct for more general tasks—and then processes the response. The researchers created the ToolScale dataset with thousands of examples from fields such as finance, sports, and medicine, where the model learns to coordinate up to 50 rounds of interactions. The result? On τ²-Bench, it achieved 80.2%, while calling GPT-5 in only 40% of cases, yet still performed better than GPT-5 alone.

Customization for everyone

Users can set their preferences—for example, prioritizing local search over internet search for privacy reasons. Orchestrator adapts accordingly, making it flexible even with unfamiliar tools such as Claude Opus 4.1 or DeepSeekMath-7b-Instruct. The entire system is designed to be fast and inexpensive, with latency measured in minutes and costs in cents. You can read the detailed report at arxiv.org.

Advertisement

Content created with help from UpTier.

SEO and GEO on autopilot. UpTier’s multi-agent systems write and optimize content for search engines and AI answers.

Discover UpTier ↗

Category:AI
Did you enjoy this article?
Discover more interesting posts on our blog
Back to blog

Related posts

OpenAI cracks down on child abuse. ChatGPT will block accounts for a full dayOpenAI cracks down on child abuse. ChatGPT will block accounts for a full day
OpenAI is strengthening child protection: a new system can detect previously unknown AI-generated images, and ChatGPT will block accounts for 24 hours if users try to obtain them.
4 min read
18. 11. 2026
Google makes Tasks free for Gemini 4 Hydrogen usersGoogle makes Tasks free for Gemini 4 Hydrogen users
Tasks is becoming available for free to hundreds of millions of Gemini 4 Hydrogen users. Google says it can call businesses, book appointments and complete purchases on their behalf.
1 min read
18. 11. 2026
OpenAI introduces GPT-6.2 Astra and Sol for professional and everyday workOpenAI introduces GPT-6.2 Astra and Sol for professional and everyday work
OpenAI has introduced two models: Astra for complex professional tasks and Sol for everyday work. Both are available in ChatGPT Work, Codex and through the API.
3 min read
17. 11. 2026
Přihlaste se k odběru našeho newsletteru
Zůstaňte informováni o nejnovějších příspěvcích, exkluzivních nabídkách, a aktualizacích.
CodedTrip

Operated by CodedTrip LLC, USA.

YouTube
TikTok