QwQ-32B: Embracing the Power of Reinforcement Learning

Original Article ↗ Hacker News Discussion ↗

Model architecture & positioning

  • QwQ-32B is repeatedly compared to DeepSeek-R1 and o1/o3-mini: seen as a focused reasoning model (math/code) rather than a broad world-knowledge system.
  • Several comments clarify MoE (mixture-of-experts): experts live inside layers, and a router picks a subset per token per layer; total active parameters can be comparable to a dense 30–40B model.
  • Some speculate MoE mainly helps for long-tail knowledge; for math/code you may only need a subset of “experts,” so a dense 32B focused on those domains can match a much larger MoE.
  • Others doubt the “experts specialize by domain” story and suggest MoE may be a temporary local optimum, with future work distilling many experts into smaller “jack-of-all-trades” dense models.

Performance and behavior

  • Many users are impressed: QwQ-32B feels “insanely” strong for its size, often close to DeepSeek-R1 and occasionally beating R1/4o on specific math/engineering questions.
  • Some warn not to trust benchmarks alone and report mixed real-world results: good but not obviously superior in all cases.
  • The model’s chain-of-thought is described as very long, self-correcting (“wait… alternatively…”), sometimes looping or “overthinking” trivial tasks.
  • A few difficult puzzles that stumped other reasoning models were eventually solved by QwQ after extended deliberation, which users found notable.

Chain-of-thought & context issues

  • Very long CoT can cause “catastrophic forgetting” where the model loses the original task or ends at the </think> tag without giving an answer.
  • Many such failures are traced to tooling defaults (e.g., Ollama silently truncating to 2k context unless num_ctx is increased), not the raw model limit (~131k).
  • Long context still degrades quality after ~20–30k tokens; commenters argue current models in general are weak at long-context reasoning.
  • Suggestions include forcing a maximum thinking budget or using structured generation to cap thinking tokens.

Running locally & hardware needs

  • Widely available via Qwen’s own chat, HuggingFace Spaces, Groq, Ollama, MLX, vLLM, etc., though some frontends have sign-in friction or misconfiguration.
  • Reports: ~20–22 GB for 4-bit quant; ~40 GB+ VRAM for higher-precision with moderate context; runs (slowly) on 32–48 GB Apple Silicon and fast on 24 GB RTX-class GPUs once loaded.
  • vLLM/TGI are reported 2–6x faster than Ollama; state of local-inference tooling is described as error-prone and under-tested (wrong chat templates, misleading context handling).
  • People share concrete Ollama tips (modelfiles, num_ctx, environment variables) and note new MLX quants for Macs.

Economics, open models & GPUs

  • Several see QwQ-32B as accelerating the “race to zero”: small, free/open models rivaling or undercutting closed frontier models; some predict trouble for companies that over-bought GPUs.
  • Others invoke Jevons paradox: more efficient models will be scaled up and used for more ambitious workloads (multi-agent systems, world models, continuous self-play), so demand for compute and NVIDIA’s position likely remain strong.
  • Some note that small, capable models favor edge devices (phones, PCs, robots), potentially helping hardware vendors like Apple and Qualcomm.

Geopolitics and national strategies

  • Thread branches into US–China–India discussion: claims that China’s strategy is to pair open-source software with robotics/industrial capacity; counterarguments say firms are profit-driven, not centrally controlled, though governments can align incentives.
  • Long subthread on US tariffs and protectionism: debate over whether tariffs actually create jobs, impact exports, and how they interact with AI/automation and the working class.
  • India is lamented as “not in the race” despite talent; another commenter notes its late economic development and past IMF/World Bank-driven reforms.

Safety, censorship & bias

  • Some celebrate QwQ as “less censored” and thus more enterprise-friendly; others strongly disagree, showing that it refuses to discuss China-sensitive topics like Tiananmen Square.
  • Internal CoT in such cases explicitly reasons about 1989 events and then decides to avoid them to comply with guidelines, which some find politically revealing.
  • Comparisons are drawn to other models (e.g., ChatGPT) that also suppress answers on politically sensitive or legally fraught topics.

User experience & ecosystem

  • Qwen’s own chat interface is praised for stability and clear per-model descriptions (including context limits and use cases).
  • There’s enthusiasm about increasingly powerful small models making local, privacy-preserving use practical, even on consumer hardware.
  • Some users still prefer commercial models like Claude for speed and polish, using QwQ as a “second opinion” reasoning engine.