Qwen3-Max-Thinking

Original Article ↗ Hacker News Discussion ↗

Capabilities and Benchmarks

  • Qwen3-Max-Thinking is seen as competitive with frontier models but not clearly ahead of Claude Opus 4.5 or GPT‑5.2, especially in agentic coding where Opus still leads on SWE-verified tasks.
  • In the shared benchmark table, Qwen shines in:
    • Instruction following / alignment (especially ArenaHard v2)
    • Agentic search (HLE with tools)
  • It lags or is middling in:
    • Agentic coding (SWE Verified)
    • Several tool-use benchmarks (Tau², BFCL, Vita, Deep Planning).
  • Some argue benchmarks are increasingly detached from day‑to‑day usefulness; others still treat them as a valuable but incomplete signal.

Open vs Closed & Local Deployment

  • Qwen “Max” models remain closed-weight; access is via Alibaba’s API only, which many see as a dealbreaker versus open-weight GLM/Minimax/DeepSeek.
  • Several users confirm there is still no open-weight model that matches top-tier hosted coders on a consumer machine (e.g., M3 Pro with 18GB RAM).
  • Best current local options mentioned: Qwen3‑coder 30B, GLM‑4.7 Flash, some quantized variants on high‑VRAM GPUs—good but clearly below Codex/Opus/GPT in quality and speed.

Pricing and Market Dynamics

  • Qwen/Alibaba pricing is unclear; no obvious subscription comparable to Anthropic/OpenAI.
  • Within mainland China, Alibaba’s models are significantly cheaper; commenters attribute this to:
    • Domestic price wars
    • Lower local cost structures
    • Direct government subsidies and “compute vouchers.”
  • Some complain Alibaba Cloud onboarding and billing (especially for reasoning tokens) make margin modeling hard.

Chinese vs Western AI Development

  • Several posts repeat the claim that Chinese frontier models trail US models by ~6–9 months.
  • A common narrative: Chinese labs heavily distill and SFT on outputs from US models due to compute constraints—keeping them close but not leading.
  • Others note that “capabilities are spiky”: with different RL focus, Chinese models could become best-in-class on specific tasks even if worse overall.
  • Debate over China’s long‑term compute advantage (energy capacity vs lagging GPU/CPU ecosystem) remains unresolved.

Censorship, Safety, and Trust

  • Qwen3-Max on Alibaba’s chat site refuses to answer questions about Tiananmen, Taiwan’s status, Xinjiang, etc., with “content security” errors; similar filtering appears in some open-weight Qwen variants’ thought traces.
  • Some see this as disqualifying for factual or research use; others shrug because they only care about coding.
  • Many draw parallels to Western models’ guardrails (drugs, hate speech, Gaza/Israel, certain individuals like a defamed law professor) and a US executive order on “woke AI.”
  • There is extended argument over whether government-mandated censorship (China) is categorically worse than corporate/soft censorship (US/EU), with no consensus.

Reasoning, Token Economics, and AGI

  • Qwen3-Max-Thinking explicitly exposes “thought” steps and is significantly slower; users speculate it consumes many more tokens per query.
  • Several point out that “better reasoning” is often just “spending more tokens,” i.e., economic tradeoff rather than pure architectural gain.
  • Concern: opaque, auto‑decided “thinking time” destroys predictable unit economics; others note newer APIs let you cap thinking effort.
  • Discussion on AGI: if strong reasoning requires huge per‑query compute, even a breakthrough model might be bottlenecked by inference capacity.

Search, Data, and the Chinese Internet

  • Qwen’s strong performance on tool‑augmented/“with search” benchmarks prompts speculation that Chinese web content or search infrastructure could be higher‑quality for certain tasks.
  • Others argue a simpler explanation: better retrieval and tool orchestration, not a fundamentally “better internet.”
  • Users dissatisfied with Western deep‑research features say they often surface low‑quality, repetitive web content; some prefer academic‑only search filters.

Developer Experience & Anecdotes

  • One user reports Qwen3‑coder significantly outperforming prior Gemini and Claude versions on complex Rust refactors (shared memory, SIMD) but at high Alibaba API cost due to large contexts.
  • Others find Qwen3-Max-Thinking slow and possibly overloaded at launch.
  • There is ongoing skepticism about “benchmaxxing” vs real‑world coding performance, but also clear enthusiasm for Qwen/GLM/Minimax as serious, closing‑gap alternatives to US incumbents.