在官方与 Artificial Analysis、LMArena 都还没有分数的时候,Stealth Models(StealthMark)给出了 Qwen4-Max-Preview 的首个独立测量。它的结论是一升一降:视觉智能高于上一代 Qwen3.8-Max,通用智能反而更低。这个组合初看矛盾,拆开看却很有信息量。
视觉维度超过上一代,与其他实测的观察互相印证:SVG 图形、3D 场景、界面生成普遍获得好评(见本站 techgogogo 六题实测)。如果这代模型的训练资源明显向视觉与生成倾斜,视觉分上涨是合理结果。
通用智能的回落主要由两个游戏项目拉低。这点值得玩味:其他测试者在游戏生成上给出的评价相当正面(一次生成可玩的太空射击、抓娃娃机)。同一类任务、不同量表得出相反结论,恰恰说明早期小样本测量的噪声有多大——量表怎么设计、采样几次、用什么提示词,都足以翻转名次。
必须强调:StealthMark 的分数是编辑估算,样本量小,不是标准化基准。它适合用来提出假设(例如「这代可能更偏视觉」),不适合用来下结论或做排名。真正的通用智能水平,要等 Artificial Analysis、LMArena 等机构的正式收录,或者更多可复现的公开测试。
参考来源:本文为本站原创改写与整理,事实与数据核对自 Stealth Models (StealthMark)(2026-11-11)
← 返回 Qwen4-Max 专题在官方與 Artificial Analysis、LMArena 都還沒有分數的時候,Stealth Models(StealthMark)給出了 Qwen4-Max-Preview 的首個獨立測量。它的結論是一升一降:視覺智能高於上一代 Qwen3.8-Max,通用智能反而更低。這個組合初看矛盾,拆開看卻很有信息量。
視覺維度超過上一代,與其他實測的觀察互相印證:SVG 圖形、3D 場景、界面生成普遍獲得好評(見本站 techgogogo 六題實測)。如果這代模型的訓練資源明顯向視覺與生成傾斜,視覺分上漲是合理結果。
通用智能的回落主要由兩個遊戲項目拉低。這點值得玩味:其他測試者在遊戲生成上給出的評價相當正面(一次生成可玩的太空射擊、抓娃娃機)。同一類任務、不同量表得出相反結論,恰恰說明早期小樣本測量的噪聲有多大——量表怎麼設計、採樣幾次、用什麼提示詞,都足以翻轉名次。
必須強調:StealthMark 的分數是編輯估算,樣本量小,不是標準化基準。它適合用來提出假設(例如「這代可能更偏視覺」),不適合用來下結論或做排名。真正的通用智能水平,要等 Artificial Analysis、LMArena 等機構的正式收錄,或者更多可復現的公開測試。
參考來源:本文爲本站原創改寫與整理,事實與數據覈對自 Stealth Models (StealthMark)(2026-11-11)
← 返回 Qwen4-Max 專題While officials, Artificial Analysis and LMArena still had nothing, Stealth Models (StealthMark) published the first independent measurement of Qwen4-Max-Preview. Its verdict splits in two: visual intelligence above the previous Qwen3.8-Max, general intelligence below it. The combination looks contradictory; unpacked, it is informative.
A visual lead over the previous generation corroborates what hands-on testers saw: SVG art, 3D scenes and UI generation drew consistent praise (see our techgogogo six-task write-up). If this generation's training budget tilted toward vision and generation, a higher visual score is the expected outcome.
The general-intelligence dip was driven mainly by two game projects — which is exactly where other testers were most positive (one-shot playable space shooters and claw machines). Same task family, opposite conclusions under different rubrics: a perfect illustration of early-measurement noise. Rubric design, sample count and prompt choice can each flip the ranking.
Be clear about what this is: editorial estimates on a small sample, not a standardised benchmark. It is good for forming hypotheses ("this generation may lean visual") and useless for verdicts or rankings. The real general-intelligence picture waits for formal listings from Artificial Analysis or LMArena, or for more reproducible public tests.
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against Stealth Models (StealthMark)(2026-11-11)
← Back to the Qwen4-Max topic page