擴散式LLM評測：NVIDIA TwoTower生成快2.4倍免重訓 | Diffusion LLM TwoTower Review: NVIDIA's 2.4x Speed Boost

By Kit 小克 | AI Tool Observer | 2026-07-31

🇹🇼 擴散式LLM評測：NVIDIA TwoTower生成快2.4倍免重訓

NVIDIA最新推出的開源擴散式語言模型（Diffusion LLM）「Nemotron-Labs-TwoTower」在7月初悄悄上線，主打不用重新訓練整個模型，就能讓生成速度提升2.42倍，同時維持98.7%的基準測試品質。這是繼Inception Labs的Mercury、Google的Gemini Diffusion之後，擴散式LLM陣營再添一位重量級玩家，也讓「LLM不一定要一個字一個字生成」這件事，開始從研究論文走向真實可用的開源權重。

什麼是「雙塔」架構？

傳統自迴歸（Autoregressive, AR）LLM像人打字一樣，一次只吐一個token，吐完才看得到下一個字要生成什麼。TwoTower的做法是把工作拆成兩座「塔」：

上下文塔（Context Tower）：凍結不訓練，負責讀懂已經生成的內容與prompt，沿用原本的自迴歸方式運作
去噪塔（Denoiser Tower）：唯一需要訓練的部分，透過遮罩擴散（mask diffusion）技術，一次性把一整個區塊的token同時生成出來，而非逐字吐出

這個架構是站在NVIDIA自家開源模型Nemotron-3-Nano-30B-A3B（混合Mamba-2、注意力機制與MoE的30B參數模型）之上改造而成。因為上下文塔是凍結的，團隊只需要訓練去噪塔，用了約2.1兆token，遠低於原始骨幹模型25兆token的預訓練量——這也是這次技術的關鍵賣點：用相對小的訓練成本，換來近2.5倍的推理吞吐量。

實際部署門檻：不是隨便一張顯卡能跑

老實說，這不是那種「筆電也能跑」的輕量模型。30B參數規模，即使有MoE架構省下部分算力，實際部署仍需要有一定規模的GPU資源，比較適合企業或研究團隊在自有機房、雲端GPU上測試，一般開發者想在本機把玩門檻不低。

另外98.7%的品質保留率是「聚合基準分數」的平均值，不代表在每一項任務上都跟原本的AR模型一樣準。如果你的應用高度依賴精確的長邏輯鏈推理，建議自己拿實際任務測過再上線，不要只看官方數字就下決定。

為什麼值得關注

擴散式LLM這條路線過去被視為AR模型的「另類玩具」，但NVIDIA這次證明：你不必從頭訓練一個全新的擴散模型，而是可以把現成的自迴歸骨幹凍結、外掛一個訓練成本低很多的去噪塔，就拿到明顯的速度提升。這種「省訓練成本換推理速度」的思路，對於想壓低GPU帳單的團隊會有吸引力。模型權重已經在Hugging Face開源，採用NVIDIA Nemotron Open Model License，想自己動手測的人可以直接下載。

好不好用，試了才知道。

🇺🇸 Diffusion LLM TwoTower Review: NVIDIA's 2.4x Speed Boost

NVIDIA quietly released an open-weight diffusion language model called Nemotron-Labs-TwoTower in early July, claiming 2.42x faster generation throughput while keeping 98.7% of the original benchmark quality — without retraining the whole model from scratch. It joins Inception Labs' Mercury and Google's Gemini Diffusion as the latest serious entrant in the diffusion LLM race, and it's a sign that generating tokens one at a time is no longer the only game in town.

How the Two-Tower Design Works

Standard autoregressive (AR) LLMs generate text like someone typing one letter at a time — each token depends on seeing the previous one. TwoTower splits the job across two towers instead:

The Context Tower: frozen, untrained, handles reading the prompt and previously committed tokens using standard causal attention
The Denoiser Tower: the only trainable part, uses mask diffusion to generate an entire block of tokens at once instead of one by one

TwoTower is built on top of NVIDIA's own open-weight Nemotron-3-Nano-30B-A3B, a 30B-parameter hybrid model mixing Mamba-2, attention, and mixture-of-experts layers. Because the context tower stays frozen, NVIDIA only had to train the denoiser — on roughly 2.1 trillion tokens, a fraction of the 25 trillion tokens used to pretrain the original backbone. That is the real pitch here: a relatively cheap training run buys nearly 2.5x more inference throughput.

The Deployment Reality Check

This is not a run-on-your-laptop model. At 30B parameters, even with MoE efficiency, you still need serious GPU capacity to run it — this is aimed at teams with cloud GPU budgets or in-house clusters, not hobbyists on a single consumer card.

Also worth flagging: that 98.7% quality figure is an aggregate benchmark average, not a per-task guarantee. If your use case leans heavily on long, precise reasoning chains, test it against your actual workload before trusting the headline number.

Why It Matters

Diffusion LLMs used to be treated as an AR side-project. NVIDIA's release shows you do not need to train a diffusion model from zero — you can freeze an existing AR backbone, bolt on a much cheaper-to-train denoiser tower, and get a meaningful speed win. That spend-less-on-training, save-more-on-inference trade-off will appeal to teams watching their GPU bills. The weights are already up on Hugging Face under the NVIDIA Nemotron Open Model License, so anyone curious can pull it down and benchmark it themselves.

好不好用，試了才知道 — the only way to know if it is actually useful is to try it yourself.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士