GLM-5.3 正式發布：傳聞屬實 — Z.ai 的後訓練編碼與網路防禦旗艦

The leak was real, the launch is here, and GLM-5.3 now has an independent score to argue about. Artificial Analysis' Intelligence Index — measured by the lab, not by Z.ai — puts GLM-5.3 at 60, tied with Kimi K3 for the top open-weights score on the board and 7 points clear of GLM-5.2's 53. The API went live this week at the same price as its predecessor, the open weights are confirmed for Friday, August 28, and the coding and cyber-defense claims Z.ai has been making since the August 14 announcement are starting to become testable. This page first tracked GLM-5.3 from its August 3 leak traces; this is the launch report, updated in place with what the launch, the API, and the first independent benchmark actually confirmed.

從洩漏到發布

The four traces that surfaced on August 3 — a "ZCode for GLM-5.3" harness page, an official docs page reachable for roughly an hour, a Bing index entry reading "GLM-5.3 Official Harness," and a commit adding a "glm-5.3" entry with JSON Schema support to Zhipu's official Java SDK — all pointed at a real, named release. Z.ai co-founder Tang Jie's "sooooooon" reply and the "epic-level plus" framing are now confirmed by an actual product rather than a rumor. The "roughly a week" timing signal this page tested — posted on X by @teortaxesTex after DeepSeek V4 Pro shipped on August 13 — held to the day: Z.ai formally announced GLM-5.3 on August 14 under the slogan "Built to Code. Ready for Cyber Defense." The follow-on came this week: on August 19 Z.ai said the GLM-5.3 API was live and open for calls, priced the same as GLM-5.2, with the model already wired into ZCode, AutoClaw, and the GLM Coding Plan.

「事後訓練」究竟帶來了什麼

架構方面的說法最值得釐清，因為這次洩漏最聳動的具體細節——參數突破一兆——是錯的。GLM-5.3 並不是更大的模型。Z.ai 表示，它沿用與 GLM-5.2 完全相同的 743B 混合專家（MoE）基礎模型（每個 token 約有 400 億個活躍參數），維持相同的 100 萬 token 上下文視窗與約 128K 的最大輸出長度，所有提升都來自規模更大的後訓練：更多長程任務環境、更多環境類型，以及更長的訓練時程；這些建立在 IndexShare 長上下文、SAO 非同步強化學習，以及已催生 GLM-5.2 的開源 slime 框架之上。「不重新訓練、全靠後訓練」這個說法是 Z.ai 自己的主張，且尚未經過獨立稽核。規模則是另一個重點：GLM-5.3 總參數達 743B，但每個 token 僅約 40B 活躍參數，輕量到可以在規模適中的叢集上自行部署，也便宜到可以大量提供服務——這正是發布當天評論所附上的「更小、更便宜、開放」定位，也與它所被拿來比較的封閉前沿旗艦模型形成強烈對比。

編寫程式與代理：Z.ai 宣稱的數字

在程式碼方面，Z.ai 報告——這些都是廠商自行通報、尚未被獨立重現的數據——Terminal-Bench 3.0 從 4.6 上升到 28.3，Z.ai 稱這是該測試基準上開放權重模型的最高分；DeepSWE v1.1 從 46.2 上升到 66.9；SWE-Marathon 大致翻倍，從 19.4 來到 42.5；Agents' Last Exam (CLI) 則從 23.8 上升到 28.5。在 Z.ai 的內部程式碼評測中，GLM-5.3 在高強度設定下以每個任務約 50K 輸出 token 取得 31.4% 的分數，而 Claude Opus 4.8 以約 120K token 取得 29.5%——Z.ai 想強調的重點是，GLM-5.3 在花費遠少於對手的輸出 token 情況下，達到了可相比擬的結果。Claude Fable 5 在最大強度下仍以 39.5% 領先該內部評測，而 Z.ai 也承認 GLM-5.3 在幾項較困難的程式碼評測上仍落後於 GPT-5.6 Sol 和 Claude Fable 5。在獨立基準重現這些結果之前，請將以上數字全都視為廠商數據。

網路防禦：無人預見的能力

這些網路安全數字才是真正的新聞，而且它們完全是由廠商自行通報的。在 CyberGym（一個白箱漏洞發現與驗證基準測試）上，Z.ai 報告 GLM-5.3 達到 84.5%，高於 GLM-5.2 的 77.2%，也領先 Anthropic 的 Mythos 5（83.8%）和 GPT-5.6 Sol（83.6%）。在 ExploitBench 上，該基準要求同時具備根因分析和可運作的漏洞利用程式，GLM-5.3 從 24.4% 翻倍以上至 54.4%，不過 Mythos 5（78.0%）仍保持領先。在 ExploitGym 上，Z.ai 報告在 2 小時的預算內完成 105 項任務，在 6 小時內完成 130 項；相比之下，GLM-5.2 只有 29 項和 39 項——同樣落後於 Mythos 5（181 和 247）。Z.ai 將這種網路安全能力描述為規模化後期訓練的湧現特性——用該公司的話說，「能力隨著訓練規模的擴大而不斷累積」——而非刻意追求的目標。

Z.ai 為其基準測試增添了一項真實世界的實際成果：在與資安團隊的測試中，GLM-5.3 在 269 個開源專案中辨識出 2,436 個漏洞，其中 1,097 個被評為重大或高嚴重性等級；最早發現的漏洞可追溯至 1981 年，平均「存在年限」為 26.6 年。這是公告中最引人注目的數字，也是最無法獨立查證的數字。該公司已將此能力搭配「安全揭露登記冊」（Security Disclosure Ledger）以進行協調式揭露、一項「可信存取」（trusted access）計畫，將敏感的網路安全功能限制給經驗證的使用者，以及一項「開源防護盾」（Open Source Shield）計畫，持續稽核重要的開源專案。

基準 GLM-5.3 必須擊敗

GLM-5.2 is the reference point the whole story hangs on. It shipped in June 2026 as a 743B Mixture-of-Experts model with roughly 40B active parameters per token, a 1M-token context window, a 128K max output, an MIT license, and open weights on Hugging Face. Independently, Artificial Analysis' Intelligence Index puts GLM-5.2 at 53 — the highest open-weights score on the index until this week. GLM-5.3 now clears it by 7 points: Artificial Analysis measures GLM-5.3 at 60 on the same index (v4.1.1), tying Kimi K3 for the top open-weights position and landing it in the frontier band alongside closed flagships like Claude Fable 5 and GPT-5.6 Sol. That is the first independent number attached to GLM-5.3, and it is consistent with the direction — if not every detail — of Z.ai's own claims. On long-horizon coding, the OrcaRouter harness measures 77.9 on Terminal-Bench 2.1, while Z.ai's best-reported GLM-5.2 figure is 82.7, which would be the first open-weight score above 80 but is vendor-reported and unreproduced. The list price is $1.40 per million input and $4.40 per million output tokens.

The scoreboard above is the leak-era projection this page published before launch — the ">1T params (rumored)" row, the unconfirmed context and license, the projected AA index. The launch corrected the biggest cell: GLM-5.3 reuses the same 743B base as GLM-5.2, so there is no parameter jump. The context window is confirmed at 1M, and the license stays unconfirmed because the open weights have not shipped yet. The projected index cell — this page's own guess of ~57–60 — was the rare projection that came in on the nose: the real number is 60, and the open question now is what happens when that score is reproduced against the actual weights.

The capture above is the independent baseline GLM-5.3's claims are measured against. GLM-5.2 tops the open-weights leaderboard at an Artificial Analysis Intelligence Index of 53. The first test of whether GLM-5.3's post-training deltas move that number has now arrived: Artificial Analysis measures GLM-5.3 at 60 on the same index — tied with Kimi K3 for the open-weights lead, 7 points ahead of GLM-5.2, and reported by the lab as independently measured.

定價與供應情況

GLM-5.3 is priced identically to GLM-5.2: $1.40 per million input and $4.40 per million output tokens (¥8 / ¥28 in the domestic listing), with cached-input reads at $0.26 / ¥2 per million. Z.ai announced the API was open on August 19, and it is reachable through Z.ai's own API, ZCode, AutoClaw, the GLM Coding Plan, and several partner gateways. One behavior change matters for API callers: requests now require "thinking" enabled across three effort levels — low, high, and max — with no off switch, a breaking change for existing integrations.

Same price does not mean same bill. GLM-5.3 runs roughly 20% more tokens per task than GLM-5.2 did on the same workloads, which a cost-per-task reading puts at about $0.68 against GLM-5.2's $0.44 — still under Kimi K3 (about $0.84) and GPT-5.6 Sol (about $1.23). That per-task math is a derived estimate from observed token usage, not a vendor figure, but it is the number that decides whether the flat $1.40 / $4.40 rate card actually saves you money.

發布為你帶來的改變

For API callers already on GLM-5.2, the practical step is a model-name change, not a project: GLM-5.2 is OpenAI-compatible and the integration carries over, with the thinking-effort caveat above. For self-hosters, the timeline is now a date rather than a guess: Zhipu promised the weights "two weeks after release" on August 14, which lands on Friday, August 28, and the open question is whether the license stays permissive. For anyone comparing models in the DeepSeek V4 Pro, Qwen3.8-Max, Kimi K3, GPT-5.6 Sol, and Claude Fable 5 tier, GLM-5.3 is now a live, independently scored variable in that ranking instead of a rumor.

The launch-day argument around GLM-5.3 is that the coding frontier has converged: for most everyday tasks, the story goes, few users can reliably tell GPT-5.6 Sol, Claude Fable 5, Kimi K3, GLM-5.2, and Qwen3.8-Max apart. If that convergence is real, the deciding factors stop being raw capability and become price, openness, and switching cost — which is exactly the corner GLM-5.3 is staking out at $1.40 / $4.40 per million on a self-hostable 743B base with weights confirmed for August 28. Whether coding models are genuinely interchangeable is an opinion, not a benchmark; the prices, the parameter count, and the weight date are not.

On the routing side, GLM-5.3 went live on OrcaRouter on August 18, the same day Z.ai's API opened — at the first-party list price, $1.40 / $4.40 per million, passed through with zero markup. The screenshot below shows GLM-5.2's page, which is exactly the shape GLM-5.3 now has: same price, same 1M-token context, same 128K max output. Routing a slice of real traffic to GLM-5.3 with automatic failover to GLM-5.2 or another proven model is a configuration change, not a rewrite — same key, no second contract. If the new model regresses on your workload, the router falls back before a page turns, and you get a quality signal on your own traffic instead of a vendor's slide. For a model whose flagship claims are still mostly vendor-reported, that is the low-risk way to find out for yourself.

接下來要看什麼

• The weights, on Friday, August 28, and the license line on the model card — permissive MIT like GLM-5.2, or something narrower. Zhipu's cyber-safety hardening is the stated reason for the two-week delay, and the "trusted access" program suggests some functions will be gated regardless.

• Whether the cyber claims hold up outside Z.ai's own harness. The 2,436-vulnerability real-world claim and the CyberGym lead are the numbers independent labs will probe first; the AA Intelligence Index measures general capability, not security.

• Where the index lands once the weights are out. The 60 is scored against the served API; the self-hosted version, with a license attached, is the one teams will actually redeploy.

• DeepSeek V4 Flash's announced price increase, which sets the pricing envelope GLM-5.3 is being judged against, and GPT-5.6 Sol's one-point lead at 61.

常見問題

GLM-5.3 是比 GLM-5.2 更大的模型嗎？

不。Z.ai 表示，GLM-5.3 使用了與 GLM-5.2 相同的 7,430 億參數 Mixture-of-Experts 基底模型，具備相同的 100 萬 token 上下文視窗，每個 token 約有 400 億個啟用參數。所有公布的效能提升都來自擴充的後期訓練，而非參數增加——這直接糾正了外洩時期關於基底模型超過一兆參數的傳聞。該架構說法是 Z.ai 自己的主張，尚未經過獨立審核。

GLM-5.3的開放權重何時會發布？

Friday, August 28. Zhipu promised the weights "two weeks after release" when it announced GLM-5.3 on August 14, and said "next Friday" when the API went live on August 19 — both readings land on the same date. The license has not been confirmed, and Z.ai has said sensitive cyber functions will be restricted to a verified-user "trusted access" program.

我應該如何處理基準數字？

Split the list. The coding jumps — Terminal-Bench 3.0 at 28.3, DeepSWE v1.1 at 66.9, SWE-Marathon at 42.5 — and the cyber results — CyberGym 84.5%, ExploitBench 54.4% — all come from Z.ai's own announcement and remain vendor-reported until an independent harness reproduces them. The Artificial Analysis Intelligence Index of 60 is the first independent measurement, and it is the number to weigh against everything Z.ai claims.

The leak was real, and the launch confirmed the name, the framing, and the timing — while correcting the one specific the rumor mill got loudest about. GLM-5.3 is the same base, post-trained hard, and now it carries an independent score to hold its vendor claims against: 60 on the Artificial Analysis Intelligence Index, tied with Kimi K3, seven ahead of GLM-5.2. The coding and cyber numbers are still Z.ai's own, the weights land on August 28, and the license line and a genuinely independent probe of the security claims are what's left to settle. Until then, the low-risk way to form your own view is a slice of real traffic and a failover to something proven.