Quasar 438B: Europe's Leading AI Model
Europe’s new Quasar 438B language model is promoted as the region’s leading AI, but many question its technical merit, noting it appears to be a compressed derivative of China’s GLM-5.2 that underperforms smaller open models like Qwen3.8-27B. Commenters debate whether it matters where open-weight models come from, weighing digital sovereignty, censorship, and the risk of hidden backdoors against practical performance and openness. The conversation also highlights skepticism toward the vendor’s quantum-computing marketing and EU funding culture, alongside broader worries that Europe is falling behind the US and China in both AI models and hardware.
Model performance vs. size and benchmarks
- Many are struck that a 438B-parameter model scores below smaller models like Qwen3.8-27B; some question whether benchmarks are being gamed.
- Training data quality and quantity are seen as more important than architecture tweaks; Qwen’s improvement from 2.5 to 3 is cited as data-driven.
- Closed-weight “frontier” models may benefit from proprietary “harnesses” (e.g., automatic retries, de-stucking prompts), while open models are benchmarked “naked.” One commenter claims a ~10% SWE-Pro boost by adding such a proxy, raising doubts about comparability.
European vs. foreign models and digital sovereignty
- Some Europeans say they don’t care where open weights come from; others argue origin matters for geopolitics, digital sovereignty, and resilience if US/China cut access.
- There is strong desire for competitive European models so locally hosted options don’t overwhelmingly depend on Chinese stacks (Qwen, DeepSeek, GLM).
- EU AI Act–style constraints and data-protection rules may require Europe-hosted, compliant models for personal data, creating a real (if niche) market.
Trust, ideology, and censorship
- Several argue LLMs inherently encode values and ideology; others wish for “non-ideological” models, which is contested as impossible.
- Concerns include backdoors, covert exfiltration, and subtle manipulation in code or outputs, especially from adversarial states.
- People expect all regions (China, EU, US) to impose their own censorship/guardrails; some prefer model diversity to balance biases.
Origin and nature of Quasar 438B
- Multiple commenters believe Quasar 438B is based on the Chinese GLM-5.2 model, possibly compressed/modified.
- External model trackers are cited as labeling it “based on GLM-5.2.” Documentation from the company’s own platform allegedly claims “capabilities identical to GLM-5.2,” with similar effort modes and text-only support.
- This fuels criticism that “Europe’s leading model” is effectively a rebadged Chinese model that still underperforms Chinese originals.
Company credibility and “quantum” marketing
- The firm’s heavy emphasis on “quantum algorithms,” “quantum AI,” and extreme compression claims (80–95% parameter reduction with minimal loss) is widely mocked as technobabble.
- Some see the effort as grant-chasing and rebadging rather than deep technical innovation; others defend the company as serious deep-tech funded through rigorous EU programs.
- Overall tone mixes hope for European success with skepticism toward the specific marketing and transparency around Quasar 438B.