一張生成的標題卡寫著「Hy4 flash vs GLM-5.3-Flash」,副標題為「較便宜的模型是你無法下載的那個」,價格行寫著「每百萬個 token 為 $0.070 / $0.281 / $0.003,對比 $0.15 / $0.50 / $0.03」,頁腳寫著「Hy4 flash 於 2026 年 11 月 16 日發布;GLM-5.3-Flash 於 2026 年 8 月 26 日開源。」
Guides & Insights

Hy4 flash 對決 GLM-5.3-Flash:更便宜的模型,正是你無法下載的那一款

作者

jinhao.song

發佈日期

最新模型 · 20查看全部模型 →
基準測試:Artificial Analysis · 每日更新
返回全部文章

Hy4 flash and GLM-5.3-Flash are aimed at the same buyer — someone who wants a million-token context window, image input and agent-grade tool calling without paying flagship rates — and they answer with almost identical price cards. Hy4 flash, released by T​encent on 16 November 2026, bills $0.070 per million input tokens and $0.281 per million output with $0.003 cached. GLM-5.3-Flash, which Z​hipu launched and open-sourced on 26 August 2026, bills $0.15, $0.50 and $0.03 at list. T​encent is roughly half the price on input and a little over half on output, and ten times cheaper on cached input, which is the line that decides most agent budgets. It is also the model with no weights, no licence, no model card and no independent benchmark, against one that is MIT-licensed on Hugging Face and scores 27.8 on the Artificial Analysis Intelligence Index.

這就是整段比較的一句話總結,而它無法乾淨地得出結論。較便宜但無法驗證,對上較昂貴但可檢查,這是真實的取捨,不是行銷問題;哪一邊勝出,取決於兩家供應商都不會告訴你的事。以下說明雙方實際公布了什麼、價差換來什麼又付出什麼代價,以及決策的關鍵轉折在哪裡。

兩張價格卡,逐行閱讀

先從已公布的數字開始,因為它們是這項比較中,兩家公司都明確陳述、且任何人都能對照帳單查核的唯一數據。

• 輸入,每百萬個 Token — Hy4 flash $0.070;GLM-5.3-Flash $0.15。

• 輸出,每百萬個 token — Hy4 flash $0.281;GLM-5.3-Flash $0.50。

• 快取輸入,每百萬個權杖 — Hy4 flash $0.003;GLM-5.3-Flash $0.03。

• 上下文視窗 — Hy4 flash 1,048,576 個 token;GLM-5.3-Flash 1,000,000 個 token。

• 每個 token 的活躍參數 — Hy4 flash 總參數 224B 中的 12B,包含 64B 的 Engram 參數;GLM-5.3-Flash 在 45 層中總參數 320B 中的 18B。

• 權重 — Hy4 flash 以計畫形式公布,未定日期;GLM-5.3-Flash 自發布日起以 MIT 授權在 Hugging Face 上提供。

獨立智能評分——Hy4 flash 未公布任何數據;GLM-5.3-Flash 在 Artificial Analysis Intelligence Index 上為 27.8,而在同一實驗室的 Coding Index 上則為 71.5。

The cached-input line deserves more attention than it usually gets. At $0.003 per million, Hy4 flash caches cheaper than GLM-5.3-Flash by a factor of ten, and cheaper than its own flagship T​encent HY4 by a factor of a hundred. For an agent that re-sends a fixed system prompt, a tool schema and a retrieved document on every step, the cache rate is the rate you actually pay most of the time, and a tenfold difference there outweighs the twofold difference on fresh input. If your workload is retrieval-heavy, Hy4 flash's advantage is larger than the headline price suggests. If your workload is one-shot generation with no reuse, the advantage narrows to the input and output ratios.

Both price cards are promotional in part. T​encent Cloud's TokenHub is running 50% off for an unstated limited period and WorkBuddy is free until 30 November. Z​hipu's 50%-off launch rate expired on 9 September 2026 and was not extended, which is why the $0.075 and $0.25 figures that circulated in late August no longer apply — the rate above is the standing one. Assume T​encent's promotion ends the same way and budget the list rate.

參數差距小於價格差距

A generated comparison card titled 'Two cheap models, two designs' with a left column for Hy4 flash reading 224B total, 12B active, 64B Engram parameters, 1M context, $0.070 / $0.281 / $0.003, no weights published, and a right column for GLM-5.3-Flash reading 320B total, 18B active, 45 layers, 1M context, $0.15 / $0.50 / $0.03, MIT weights on Hugging Face. A footer reads 'Hy4 flash figures are Tencent's, published 16 November 2026; GLM-5.3-Flash price and licence per Zhipu, 26 August 2026.'

Hy4 flash activates 12 billion parameters per token against GLM-5.3-Flash's 18 billion, a third less compute per token, and that is the mechanical reason it can be sold at half the price. The interesting part is the total: 224 billion against 320 billion, with 64 billion of T​encent's total sitting in Engram parameters that hold learned associations rather than running in every forward pass the way a routed expert does. T​encent has not published a layer count, expert count or routing width for Hy4 flash, so the closest thing to an architectural comparison is that both are sparse mixture-of-experts models with a small active fraction and both vendors chose that design for the same reason — serving cost scales with active parameters, not total ones.

GLM-5.3-Flash's design is documented in more detail, and that documentation is checkable because the weights are public: 45 layers, hybrid sparse-plus-linear attention, an approach to residual connections Z​hipu calls Manifold-Constrained Hyper-Connections, and an IndexPool mechanism that compresses four indexer key vectors into one weighted pool to cut long-context latency. Pre-training ran on a 30-trillion-token multimodal corpus. None of that is independently audited, but it is at least a description specific enough to be falsified by anyone who downloads the checkpoint.

The asymmetry is not that Z​hipu published and T​encent did not. Both published; they published different kinds of thing. Z​hipu described a model, which is a claim about structure that a checkpoint can confirm or refute. T​encent described a result — a KV cache a sixth the size of its flagship's, output length halved by self-distillation, better than 150 tokens per second of decode from four-bit quantization and a megakernel — which is a claim about behaviour that only reproduction can settle. Results are the more commercially interesting claims and the harder ones to check. On a model with public weights, a serving claim gets tested by strangers within weeks; on a model without them, it stays a sentence on a research page for as long as the vendor wants.

其中一個你可以下載

對相當比例的買家而言,這是決定性的差異,而且差距懸殊。GLM-5.3-Flash 採用 MIT 許可證,提供 FP8 與 BF16 版本,位於 Hugging Face 上的 zai-org/GLM-5.3-Flash,且自 2026 年 8 月 26 日起便是如此。MIT 允許商業使用、修改、微調與再散布,且沒有營收門檻。該模型已廣泛公開使用近三個月,這意味著推論堆疊已能配合它:已有量化版本、已有部署方案,而任何關於其行為的說法都能透過實際運行來檢驗。

Nothing like it exists for Hy4 flash. The commitment there is a sentence about future plans, hedged with an intention to keep iterating, and it names no licence and no date. T​encent has done this properly before — the August preview landed under Apache 2.0 with weights on four hosting sites and vLLM and SGLang recipes on day one — so the pattern is not in doubt. What is in doubt is the timidity of the current language, which reads like a commitment to a process rather than to an artefact. Anyone planning to fine-tune, self-host, or build a cost model around running their own inference should treat Hy4 flash as API-only until a repository appears.

The same asymmetry applies to evidence. GLM-5.3-Flash has two numbers from a harness neither vendor wrote: 27.8 on the Artificial Analysis Intelligence Index and 71.5 on its Coding Index. Both are checkable, and the pair says something specific: a 71.5 on coding against 27.8 on the general index describes a model that is genuinely good at one thing and ordinary at everything else, which is exactly what a cheap tier should be. That is the kind of distinction a single headline score erases, and it is why the remaining vendor figures in this piece are presented as claims rather than results. Hy4 flash has no independent score, and its benchmark results — DeepSWE, Terminal-Bench 4.0, APEX-Agents, OSWorld 2.1 and T​encent's internal suites — exist only as chart images on the vendor's research page. That is not a reason to distrust them. It is a reason to treat them as hypotheses.

各廠商量測了什麼,以及它擊敗了誰

兩家公司都發表了比較,而它們各自拿自己與不同的對象相比,這說明了它們各自認為自己的弱點是什麼。

T​encent's headline is a blind test: 85 WorkBuddy users, 27 task categories, 259 real daily-work tasks, with Hy4 flash reported as landing close to Claude Haiku 5.5 on satisfaction while carrying the lowest per-task cost and shortest task duration in the panel. Note "close to," not "better than," and note the panel is the vendor's own employees in the vendor's own product. The supporting claim is per-task cost on WorkIndex, T​encent's internal eight-suite benchmark: one-thirteenth of Hy4 preview, one-twentieth of T​encent HY4, one-eighth of DeepSeek-V4.1-Flash-1017. Against GLM-5.3-Flash the only published figure is a 2.52% gain in general chat inside 元宝, which is a product metric on one consumer app and does not transfer to anything you would run.

Z​hipu's positioning is different: it sells the model as frontier-adjacent capability at flash cost, prices it against its own 743B flagship, and points to open weights and an exploit-evaluation result as proof the cheap model is not a toy. Its internal figures have already taken one independent correction, and the AA number is the one to quote.

實際後果是,兩家廠商都沒有在中立測試框架上發表正面對決結果,而你真正想要的比較——相同任務、相同提示、相同測試框架、兩個模型——並不存在。你得自己跑。你已經在做的十項任務,只要記錄 token 數量,就足以針對你的流量情境定案,而且比任一份發布頁都更有價值。

對多模態的兩種不同押注

GLM-5.3-Flash 原生支援文字、圖像與影片輸入,是首個這麼做的 GLM-5 模型,而非透過獨立適配器來處理視覺輸入。這是已有文件記載且可測試的說法,而且權重公開,因此無須徵求許可即可查核。

Hy4 flash is described as multimodal, and the WorkIndex suite it was measured on includes visual understanding as one of eight task types. T​encent has not said whether the vision path is native or an adapter, has not published the encoder's size, and has not said whether video is supported at all. The most that can be said is that image input is part of the pitch and that the model's predecessor in the family — T​encent HY4, general availability 4 November 2026 — added native vision and computer use as its headline feature, so the capability is plausible for a model distilled from it. Plausible is not documented.

如果截圖、圖表、掃描文件或影片影格在你的流程中扮演關鍵角色,GLM-5.3-Flash 就是你能在正式採用前先驗證其多模態行為的模型。如果圖像只是偶爾出現,這項差距的重要性就不及價格。

決策的轉捩點

Choose Hy4 flash when the volume is the problem and the task is office-shaped. Document processing, retrieval over a long corpus, slide generation, simple frontend work, computer-use loops — these are the workloads T​encent built it for, the ones its internal product data covers, and the ones where a tenfold cache discount and a 50% output-length cut compound into something an order of magnitude cheaper. Run it on your own tasks first and watch output tokens: if the model really is terser, your bill falls further than the rate card predicts, and if it is terser because it is shallower, you will see that in the same measurement.

Choose GLM-5.3-Flash when you need the artefact, the audit, or the modality. Fine-tuning, on-premises serving, regulated data, a licence you can show a lawyer, video input, or a score from a harness nobody at Z​hipu wrote — all of these point one way. At $0.15 and $0.50 it is not expensive; it is roughly a tenth of GLM-5.3's list price for a model that has already survived independent measurement.

當工作負載混合時,兩者都選。 誠實解讀這兩個版本,會發現它們其實並非真正的替代品:一個是為大量日常工作量而做成本最佳化的 API 模型,另一個是開放檢查點,具備已記錄的架構,適合任何需要脫離廠商控制的用途。能同時連到兩者的路由器,比起只選一個卻選錯,是更便宜的答案。

A screenshot of the OrcaRouter model page for GLM 5.3 Flash, slug z-ai/glm-5.3-flash, credited to Z.ai and dated 2026-08-26, describing a 320B-total, 18B-active native multimodal model with a 1M-token context that takes text, image and video in and returns text. The specification strip reads max output 128K, input $0.15 and output $0.50 per million tokens, p50 time to first token 3.82 seconds and p95 10.00 seconds. The pricing panel lists $0.150 input, $0.500 output and $0.030 cache read per million tokens, and the public benchmark panel shows 71.5 on the AA Coding index and 27.8 on AA Intelligence. A button reads 'Get the GLM 5.3 Flash API'.

今天實際上可呼叫的是什麼?

This is where the two models differ operationally as well as commercially. GLM-5.3-Flash is in the OrcaRouter catalogue as z-ai/glm-5.3-flash at the provider's list price of $0.15 per million input and $0.50 per million output, with a rate-limited free tier alongside it for evaluation. Hy4 flash is not routed here — T​encent has not made it available to us — so anyone comparing the two needs T​encent's own TokenHub for the flash side and can use the same OrcaRouter key for the G​LM side.

That split is worth using rather than resenting. OrcaRouter passes provider list price through with no markup, which means the G​LM side of your comparison is billed at exactly Z​hipu's rate and stays correct if Z​hipu moves it; there is no reseller table to go stale. And for the evaluation itself, having tencent/hy4-preview on the same key as z-ai/glm-5.3-flash is useful even though it is not the model in question: the preview is the model Hy4 flash was distilled from, so it is the natural control when you test whether the small model's answers hold up. One key, more than 200 models, and a failover path if a launch-week endpoint turns out to be unreliable — which, for a model released six days ago with no public uptime record, is a reasonable thing to want.

A generated decision card titled 'Pick by constraint' with three rows: high volume office-shaped work points to Hy4 flash on price per task; fine-tuning, on-premises serving or a named licence points to GLM-5.3-Flash under MIT; mixed workloads point to routing both from one key. A footer reads 'GLM-5.3-Flash scores 27.8 on the Artificial Analysis Intelligence Index; Hy4 flash has none published.'

價格不是問題

T​encent has built the cheaper model, and on the numbers it is cheaper by a margin that matters: half on input, a little over half on output, a tenth on cache. What it has not built is a model you can inspect. No weights, no licence, no architecture detail, no independent score, benchmarks behind pixels, and a headline comparison drawn from a panel of its own staff. Z​hipu built the more expensive one and published enough that its weaknesses — including a vendor index that independent measurement cut by fifteen points — are on the record where you can weigh them.

發表六天後,這個選擇的輪廓便是如此:相對於一個沒有文件說明的模型,已顯示出的折扣;或是一個有文件說明、但價格仍然非常低的模型。這個折扣很可能是真的,而文件說明的缺口很可能是暫時的。兩者都不確定,而弄清楚的方法,是花一個下午去測量,而不是花一季去選擇。

本文中的比較3

根據本文內容識別 · 基準測試:Artificial Analysis · 每日更新