Qwen3 is a fantastic open-source model

Tijdlijn van de aflevering

1. 0:00

   Qwen3 launches as an open-source model family rivaling Gemini 2.5 Pro on major benchmarks.

   • Qwen3's flagship 235B model with 22B active parameters scores 85.7 on ArenaHard versus Gemini 2.5 Pro's 92, and beats Gemini 2.5 Pro on LiveCodeBench 70.7 to 70.4 and CodeForces ELO 2056 to 2001.
   • The 30B mixture-of-experts model with only 3B active parameters is highlighted as the best of the batch, scoring 91 on ArenaHard versus GPT-4o's 85 and 62 versus 32 on LiveCodeBench.
   • Qwen3 is optimized for agents and coding, with the 235B model scoring 70.8 on BFCL function calling versus 62.9 for Gemini 2.5 Pro, and even the 32B dense model scores 70.3.
2. 2:06

   Qwen3 introduces adjustable thinking budgets and a hybrid thinking/non-thinking mode.

   • The hybrid model lets users control how many tokens the model spends reasoning, with performance increasing smoothly as more thinking tokens are allocated across benchmarks like AIME24, AIME25, LiveCodeBench, and GPQA Diamond.
   • Non-thinking mode gives quick near-instant responses for simple tasks, while thinking mode reasons step by step for complex problems, enabling task-specific budget configuration.
   • This flexibility is especially useful for vibe coding, where hard tasks like building features warrant deep thinking but routine commands like running tests and deploying do not.
3. 3:56

   Sponsor segment: Zapier's MCP server connects AI agents to thousands of apps.

   • Zapier released an MCP server giving agents access to over 7,000 apps, configured by simply selecting apps and plugging the generated URL into tools like Windsurf, Claude Desktop, or Cursor.
   • Zapier has over a decade of work automation experience and offers a free plan, making it easy to connect Qwen3 to real-world tools without writing code.

Kernbegrippen

• Qwen3— The newly launched open-source model family that is the main subject of the episode.
• open-source model— Qwen3 is fully open-source with open weights, making it accessible to everyone.
• hybrid thinking mode— Qwen3 can switch between step-by-step reasoning and quick responses, with adjustable thinking budget.

Opvallende citaten

• We now have a completely open-source and open weights model that is comparable to Gemini 2.5 Pro.

  🤯

  — Highlights the surprising fact that an open-source model can now match a leading proprietary frontier model.

  Ga naar 0:00

• Code Forces actually scored a higher ELO rating 2056 as compared to 2001 for Gemini 2.5 Pro.

  🤯

  — Reveals that Qwen3 outperforms Gemini 2.5 Pro on a competitive coding benchmark, a counterintuitive result for an open-source model.

  Ga naar 0:41

Praktische conclusies