Qwen4-Max
Qwen4-Max, Alibaba’s new flagship large language model, is being evaluated against US “frontier” models such as Claude, GPT and DeepSeek, with many noting that while its raw capability is solid, benchmark aggregates like Artificial Analysis currently place it below the very top tier and question its cost‑performance. Commenters are especially interested in the forthcoming smaller open‑weight variants (27B and Flash), since prior Qwen releases punched above their size on local hardware and in coding/agent workloads. The thread also surfaces broader concerns around benchmark gaming, distillation crackdowns by US labs, censorship and safety filters in Chinese‑hosted APIs, and the widening gap between closed and open‑weight model capabilities.
Overall reception and positioning
- Many see Qwen4-Max as an incremental but not frontier-breaking model. Several comments say it lands “one tier below” the very top US models.
- Users are more excited about upcoming Qwen4 27B / Flash open-weight variants than this closed-weight Max release; prior Qwen 3.8 Flash/27B are widely praised as “workhorse” and “SOTA-for-their-size”.
Performance, speed, and cost
- Some report Qwen 3.8 Max/Flash Next as very capable for coding, agents, and reverse engineering, often “good enough” vs Opus/Fable/Sol but slower and/or more token-hungry.
- Qwen4-Max scores poorly on Artificial Analysis’ aggregate index (intelligence 43, relatively high $/task in that setting) and roughly similar to Kimi 3.1, though other benchmarks (Vals, TerminalBench 4) place it closer to Astra-level.
- A number of practitioners emphasize that for long‑running or complex workloads, “max effort” reasoning can be faster overall despite higher per‑task cost, due to fewer corrections.
Benchmarks and Artificial Analysis controversy
- Large subthread questions AA’s methodology: frequent index rewrites, opaque weighting, private benchmarks, and perceived systematic favouring of Anthropic/OpenAI.
- Others defend AA as one of the better multi-model views available, arguing private evals are needed because labs train on public benchmarks.
- Consensus: benchmarks are heavily benchmaxxed; serious users increasingly rely on custom internal eval suites.
Open weights vs frontier labs
- Concern that open-weight Chinese models are starting to lag since US labs introduced anti-distillation measures and have more compute.
- Counterpoint: open-weight ecosystem focuses on efficiency (quantization, inference engines like Strata/Sushi) and local deployability; smaller Qwen and DeepSeek Flash models already satisfy many “assist human” tasks.
- Debate over long‑term viability of open vs closed models, with some arguing economics and compute asymmetry favour closed, others betting on local “good enough” models.
Safety, censorship, and politics
- Reports that Chinese-hosted frontends (Qwen, DeepSeek, Moonshot/Kimi) refuse or filter politically sensitive China topics, while open-weight or third‑party‑hosted variants are more permissive.
- Complaints about US labs’ safety filters (piracy, biology, security research) being overzealous and sometimes secretly modulated to block distillation.
Pelican SVG micro‑benchmark
- Thread reprises the “pelican on a bicycle” SVG test across reasoning levels. Qwen4-Max does better than 3.8 overall; only at highest reasoning does it get key geometric details right.
- Some find this a fun, light‑weight qualitative check; others dismiss it as saturated, meme‑like, and possibly in training data, though no clear evidence of explicit “pelicanmaxxing”.