Atria Dawn 與 GLM 5.2：同一個基礎，朝相反方向推進

關於這場對決，最有用的單一事實就藏在較小模型自己的規格表裡：Atria Dawn Preview是上海人工智能實驗室於9月14日直播發表的全新開放權重代理式模型，建構在744B參數的MoE GLM 5.2基礎之上——正是Z.ai在6月推出、並一直以每百萬詞元1.40／4.40美元價格作為其開放權重旗艦販售的那個模型。所以這並不是新進者試圖在建造上超越既有廠商的情況；而是同一塊矽晶，被兩個不同實驗室朝兩個方向做後訓練，然後被當成兩種不同產品來定價。在任何一列基準數據之前，這個事實就已經告訴你這場比較真正要談的是什麼：你願意為了代理式迴圈放棄多少GLM 5.2的推理能力——以及為了兩者兼得，你得付多少錢。

這裡的資料來源偏頗得相當關鍵。GLM 5.2 的獨立評測紀錄相當單薄——其開放權重上限是由 GLM 5.3 所界定，而 GLM 5.3 已於 8 月 14 日取代了它，而曾將 GLM 5.2 評為 53 分的 AA Intelligence Index 之後也已重新評分並更名。每一項有利於 Atria Dawn Preview 的基準測試，都是廠商在其模型卡上自行報告的，沒有任何中立實驗室加以重現，而其位於 api.atria-asi.ai 的國際 API 至今也尚未公布價格。在模型卡將兩者並列於同一網格比較之處，Atria Dawn Preview 在探索與工具使用方面領先（BrowseComp 92.5、DeepSearchQA 96.0、BFCL v4 77.0），卻在純編碼項目上落後於 GLM 5.2 一脈——而這些都尚無外部實驗室加以證實。

同樣的基礎，兩種後訓練

Both models are 744B-total MoEs with the same 8-expert-per-token activation pattern, because they literally start from the same weights. The difference is what each lab did after that. Z.ai took GLM 5.2 and tuned it as a general open-weights flagship: text-only, 1M context, 128K output, MIT license, with a 40B-active class profile that made it one of the cheapest frontier-grade models to serve — the thing it is still remembered for is being the highest-scoring open-weights model on the independent index before GLM 5.3 arrived a week after Intern-S2.

The Shanghai AI Laboratory took that same base and tuned it into a research-loop agent. Atria Dawn Preview's four capability pillars — Discovery, Creation, Delivery, Cybersecurity — all describe the same loop: analyse a problem, design a solution, use tools, write and run code, read the experimental result, recover from failure, iterate. The lab is explicit that the result is text-only (image and PDF input get a 400), that the context window is 256K rather than 1M, and that the weights are MIT-licensed and downloadable today in BF16 (353 shards) or FP8 (177 shards). The base is GLM 5.2's; the product is not.

• Base — both 744B-total MoE on the GLM-5.2 foundation, 8 experts active per token

• Context — Atria Dawn Preview 256K text-only vs GLM 5.2 1M, 128K output

• License — both MIT; Atria BF16 + FP8 shards, GLM 5.2 the familiar 1.5TB-class download

• Price — Atria unpublished (international API) vs GLM 5.2 $1.40/$4.40 per 1M

What the vendor-reported grid actually shows

The model card compares Atria Dawn Preview against a field that includes GLM 5.3 rather than GLM 5.2 itself, which is a small editorial choice that flatters the open-weights column. Against that field, Atria Dawn Preview's reported strengths are discovery and tool use: AutomationBench 53.8, BrowseComp 92.5, DeepSearchQA 96.0, WideSearch 81.9, BFCL v4 77.0, CyberGym 86.5. Its reported weaknesses are the coding rows where a tuned generalist usually wins: SWE-bench Pro 59.6, Terminal-Bench 2.1 78.3, JobBench 50.3. Nothing here is independently verified, and every row is the lab's own harness on the lab's own prompts — but the split is consistent enough to read as a real design choice rather than noise. GLM 5.2, for its part, is the model whose June-era independent ceiling (AA Index 53, the then-highest open-weights score) was superseded by GLM 5.3 the month before Atria shipped.

Which one you should route to

If you need a million-token open-weights generalist for long-context coding and knowledge work, GLM 5.2 at $1.40/$4.40 is a proven, MIT-licensed workhorse that has been self-hosted and production-served for a quarter — and on a gateway like OrcaRouter, one API key and zero per-request markup later, it is available the moment you need it, with automatic failover if Z.ai's endpoint ever wobbles. If your job is an open-ended research loop — a task that needs the model to keep reading, running code, and iterating until something is verifiable — Atria Dawn Preview's tuning is aimed exactly there, but you are trading the 1M window, a known price, and an independent record for a 256K preview with no published rate and no neutral scores.

誠實的結論是一個岔路口。已經依賴 GLM 5.2 進行長上下文開放式工作的團隊，應該按兵不動，並觀察獨立評測者對 Atria Dawn Preview 的報告，再把正式生產路徑切換過去。若團隊的瓶頸是迴圈本身——模型停下來詢問，而不是執行——就有充分理由試用這個新產品，最好放在通往 GLM 5.2 的容錯移轉路徑之後，這樣即使預覽版出錯，實驗也不會付出代價。相同的基礎，兩款產品；而對大多數讀者而言，決定性的一欄不是基準測試表格，而是表格從未顯示的上下文視窗與價格。