Qwen 3.8

Original Article ↗ Hacker News Discussion ↗

Alibaba’s announcement of Qwen 3.8, a 2.4-trillion-parameter large language model that is promised as open-weights, is seen as part of an escalating race among Chinese labs (Kimi, DeepSeek, GLM, etc.) to match or surpass U.S. frontier models like Anthropic’s Fable and OpenAI’s GPT line. Commenters debate the strategic motives behind China’s push for powerful open-weight models — from commoditizing “intelligence” and undercutting U.S. incumbents to genuine open-source ideals — while also weighing practical tradeoffs such as censorship, guardrails, cost, performance, and the need for smaller models that can run locally. Many view the growing availability of strong open-weight models as both a competitive threat to closed U.S. labs and a significant win for developers and smaller organizations seeking control over their AI infrastructure.

Announcement & Access

  • Qwen 3.8-Max is announced as a ~2.4T-parameter model, currently via Alibaba’s cloud/token plan and apps; open weights are promised “soon.”
  • The timing is linked by commenters to the World AI Conference in Shanghai, where multiple Chinese labs (Qwen, Kimi, others) released frontier models.
  • Some speculate it’s also a competitive response to other recent Chinese open-weight releases (e.g., Kimi K3, GLM 5.2+), though others note such models take months to train.

Scale, Capability & Comparisons

  • The size (2.4T) is seen as huge; users frame it as part of a move from “value” models to “huge, slow, very smart” Chinese models.
  • Claims in marketing that it’s “second only” to top US frontier models are treated with interest but skepticism pending hands-on testing.
  • Prior Qwen models get mixed reviews: some find 3.6 27B/35B excellent for local coding and agents; others report 3.7 Pro/Max as verbose, off-track, or weak for SWE work.
  • DeepSeek V4, GLM 5.2, Kimi K3 and US frontier models are frequent reference points; people disagree on which is “best,” but many say Chinese models are now “good enough,” especially per dollar.

Open Weights, Strategy & Geopolitics

  • Many see Chinese open-weight releases as a deliberate “commoditize the model” strategy:
    • Undercut US closed labs’ margins and valuations.
    • Drive usage of Chinese clouds, hardware, and ecosystems.
    • Build global soft power and reduce dependence on US APIs, especially within China.
  • Others emphasize intense domestic competition (“involution”): firms aggressively subsidize and open models to gain users, even at low or negative profit.
  • Government influence is debated: some cite Xi’s pro–open source AI remarks and state investment; others stress firms are still profit-seeking and see models as infrastructure to sell compute and applications.

Open vs Closed, Safety & “Deceleration”

  • Strong support for open weights:
    • Local control, no “rug pull,” no guardrail lockouts.
    • Example: a HuggingFace incident where closed APIs blocked security forensics, forcing use of a self-hosted open model.
  • A contrary view: powerful open models could reduce frontier CAPEX and slow progress (“deceleration”), which some see as good for safety, others as overstated or US-centric.
  • Concerns exist that US labs may seek regulation to hobble open-weight competitors.

Local Models & Hardware

  • Many want mid-size open variants (e.g., 30–120B dense/MoE) optimized for ~24–128GB VRAM, citing Qwen 3.6 27B, 35B MoE, Gemma 4 31B, Bonsai as current sweet spots.
  • Techniques like quantization, MoE, SSD offload, and multi-token prediction are discussed as ways to run larger models locally, albeit with speed trade-offs.
  • Some argue small capable models will continue to matter for privacy, offline use, and as cheap agents, even if giant models dominate benchmarks.

Censorship, Bias & Trust

  • Multiple comments say hosted Qwen is among the most censored Chinese models on politically sensitive topics; open weights plus “jailbreak” prompts can reduce refusals but not remove underlying training omissions or biases.
  • One camp worries about subtle ideological slant (e.g., omissions on Tiananmen, Uyghurs, Taiwan) propagating through widely adopted open models.
  • Others counter that:
    • Most users don’t use LLMs for politics.
    • Western models have their own political blind spots.
    • Open weights can be fine-tuned to counteract detected bias, whereas closed models can’t be audited or fixed by users.

Developer Experience & Use Cases

  • Experiences diverge sharply:
    • Some find Qwen excellent in coding harnesses and agent setups, especially when grounded with tools, memory, and careful prompts.
    • Others say Qwen 3.7 is unusable for SWE, preferring DeepSeek V4 Pro/Flash, GLM 5.2, or US frontier models.
  • Cost-performance:
    • DeepSeek V4 (especially Flash) is widely praised for extreme cheapness and speed, though benchmarks suggest higher hallucination rates.
    • Chinese models’ verbosity and long “thinking” traces can erode their raw token-price advantage in agentic workflows.

Meta & Cultural Tension

  • Several note that discussions of Chinese models on HN quickly shift from technical details to geopolitics, human-rights issues, and distrust of Chinese firms.
  • Some participants from China express frustration that this overshadows technical evaluation and the “geek spirit” behind the work.