Nemotron 3.5 Super VL：NVIDIA 的 120B 開源視覺模型正式推出，而 5 倍效能的宣稱至今仍無記分板可佐證

N​VIDIA N​emotron 3.5 Super VL shipped on October 15, 2026 — an open 120-billion-parameter vision-language model with open weights, open datasets and training recipes, and a claim attached to it that nobody can check yet. The company's launch post calls it "the fastest open agentic model in its class — 5x faster than the next best," and attributes the number to the Artificial Analysis Intelligence Index leaderboard. As of this writing, Artificial Analysis does not list Super VL at all; its N​emotron 3.5 entries stop at N​emotron 3.5 Lightning. That does not make the claim false. It makes it vendor-reported, and it should be read that way until a third party runs the checkpoint.

The other thing worth knowing up front is a three-day gap. The weights went up on Hugging Face at 2026-10-12T15:32Z — the BF16 post-trained build, the base model and the NVFP4 build all within a minute of each other — and the announcement post did not appear until October 15 at 15:00Z, with the model card's own release-date field still reading 10/15/2026. N​VIDIA uploaded first and announced later. That is now the normal shape of an open-weights release: the repository is the primary source, and the blog post is the commentary.

實際推出了什麼，以及什麼沒有推出

三個 Super VL 檢查點，現已全部上線：

• N​VIDIA-N​emotron-3.5-Super-VL-120B-A12B-BF16 — the full-precision post-trained model, the reference build and the one to fine-tune from.

• N​VIDIA-N​emotron-3.5-Super-VL-120B-A12B-NVFP4 — the quantized deployment build, produced with N​VIDIA Model Optimizer and quantization-aware distillation from the BF16 model.

• N​VIDIA-N​emotron-3.5-Super-VL-120B-A12B-Base-BF16 — the pre-post-training base, for anyone who wants to run their own post-training rather than adapt N​VIDIA's.

There is a fourth repository, N​VIDIA-N​emotron-3.5-Super-VL-120B-A12B-NVFP4-LiLiCorr, which is not a model you prompt. It is a speculative-decoding draft checkpoint: load it alongside the NVFP4 target and it proposes tokens ahead of the main model. Its own card dates it 10/15/2026 and describes it as intended for LiLiCorr deployments tuned for DGX Spark and low-concurrency data-center work.

Two things did not ship, and each matters. There is no price. N​VIDIA's announcement states the model is available as an N​VIDIA NIM microservice — a production microservice the company says is refreshed every two weeks with security patches and performance work — but publishes no per-token figure anywhere, for NIM or for the third-party platforms carrying it. And there is no independently measured speed number, for the reason above. Everything in the next two sections is N​VIDIA's own reporting.

用一段話說明架構，因為它能解釋部署的來龍去脈

Super VL 是一款 Mamba-2 與 Transformer 的混合式混合專家模型：交錯排列的 Mamba-2、MoE 與注意力層，每個 MoE 層有 512 個路由專家，再加上一個 C-RADIO 視覺編碼器，以及一個加裝在語言主幹上的視覺語言連接器。總參數量 120B，每個 token 啟用 12B。上下文長度最高可達 1M 個 token，且原生內建多 token 預測（Multi-Token Prediction），這正是 LiLiCorr 草稿檢查點所運用的機制。輸入為文字、圖像與影片——影片為最長兩分鐘的 MP4，以影格取樣——文件則可以頁面影像搭配文字的形式輸入。輸出僅限文字。

That shape is why the deployment table is the most useful part of the documentation. N​VIDIA's BF16 build wants 4× H100 or 2× GB200. The NVFP4 build wants 1× DGX Spark, 1× H200, 2× H100, or 1× GB200. The gap between those two lines is the whole reason the quantized checkpoint exists, and it is the difference between "a 120B model you rent" and "a 120B model on the desk."

The benchmarks N​VIDIA published, and the two problems with them

Label all of this vendor-reported. N​VIDIA evaluated Super VL through NeMo Gym and published the recipes, containers, prompts and scoring settings — which is more than most labs do — but N​VIDIA trained the model, chose the baselines and ran the harnesses. Nothing here has been reproduced outside the company.

The agentic numbers are the ones the launch leans on. SWE-bench Verified 71.10, SWE-bench Multilingual 67.72, SWE-bench Pro 51.78, tau2-bench retail 69.24, airline 76.50, telecom 73.19, BrowseComp 78.91, and GraphWalks 89.71 at the over-1M-token setting. On reasoning, GPQA-Diamond 89.52 and HLE without tools 31.79, rising to 46.01 with tools — that last figure, N​VIDIA notes, comes from a separate internal harness rather than the standard path.

On vision, which is the new capability relative to the text-only N​emotron 3.5 Lightning from August: CharXiv-RQ 80.82, MathVision 78.16, V* 85.60, OCR-Reasoning 65.08, OCRBench-v2 at 68.12 English and 59.15 Chinese, MMLongBench-Doc 62.82, OSWorld 52.14, and Video-MME-v2 32.85.

現在談兩個問題。

First, the Terminal-Bench 2.1 figure is stated twice and does not agree with itself. The model card reports 66.70. The technical report's summary table reports 60.67. Everything else in those two documents lines up, so this is not two different models — it is one number carrying two values in N​VIDIA's own paperwork, and until the recipes are re-run it is not safe to quote either. That is the sort of thing a single independent replication settles in an afternoon, which is exactly why the absence of one matters this week.

Second, the headline efficiency claim and the headline intelligence claim pull in opposite directions, and only one of them is checkable. N​VIDIA says Super VL reaches accuracy on multimodal agentic tasks "comparable with N​VIDIA's largest reasoning model at roughly one-fifth the parameter count" — N​emotron 3 Ultra 550B A55B has 55B active parameters against Super VL's 12B, a ratio of 4.6×, and the report says Super VL's 51.78 on SWE-bench Pro does surpass Ultra. Fine.

But look at the columns where Super VL is genuinely behind. MMLU-ProX 76.91, against Qwen3.5-122B-A10B at 83.2 and Qwen3.8-Flash-Next at 83.0. AA-Omniscience accuracy 26.20, which is close to the floor of that benchmark and sits below DeepSeek-V4-Flash-0731's 45.0. Video-MME-v2 32.85 against Qwen3.8-Flash-Next at 45.1. BabyVision 29.60 against 46.8. Agentic-terminal at 60.67 or 66.70 against 85.80. Super VL is not the accuracy leader in this weight class on most rows; it is a 12B-active model that N​VIDIA is selling on speed and on the breadth of what it ingests in one call, not on beating everything at everything.

效率數字，這些才是真正有趣的

把智能比較剔除後，工程數字就很具體：

• Speculative decoding — N​VIDIA's post-training MTP boosting lifts average acceptance length from 4.287 to 4.718 tokens under greedy decoding and 3.787 to 4.612 under probabilistic draft sampling.

• Quantization — the released NVFP4 checkpoint recovers 99.62% of BF16 accuracy as a median across 36 benchmarks, so the deployment build is close to lossless by N​VIDIA's own measure.

• 吞吐量 — 在單一 GB200 上，於 20 回合的代理式代理工作負載中，最高可達 4,856 個輸出 token/秒/GPU，同時維持至少 50 tokens/秒/使用者；在相同條件下，NVFP4 的吞吐量為 BF16 的 2.43 倍。

• 相對輸送量——在該工作負載上，分別為 Qwen3.5-122B-A10B 的 2.34 倍與 Qwen3.8-Flash-Next 的 1.90 倍。

• Serving stack — on N​VIDIA Dynamo, N​VIDIA claims up to 8× more concurrent users at 85% lower latency.

那些全都是廠商提供的數字。其中最容易被任何人造假的是 4,856 tokens/s/GPU 這個數字，因為它是特定 GPU 與特定工作負載的屬性，而不是在評分工具框架下執行的一次基準測試結果。

部署它，以及這兩種組建在硬體上花費你多少

Both vLLM and SGLang are named as supported runtimes, and N​VIDIA ships working launch configurations: four-way tensor parallelism with three MTP draft tokens on 4× H100 for BF16, two-way with five draft tokens on 2× GB200. Hardware coverage is Hopper and Blackwell, Linux only, and N​VIDIA lists DGX Spark, DGX Station, RTX PRO, RTX and Jetson on the client side — the same silicon spread the model is positioned against.

對於在發布日讀到這份內容的團隊來說，實際的決定是要拉取哪個檢查點。如果你正在進行微調，BF16 版本是唯一合理的起點，因為 NVFP4 版本已經過量化感知蒸餾，進一步訓練無法與它乾淨地組合。如果你是在做推論服務，NVFP4 版本才是能塞進單張 H200 或一台 DGX Spark 的版本，而 99.62% 的中位數恢復率讓它成為預設選擇，除非你已測量到特定的效能退化。只有在打算自行撰寫後訓練流程時，拉取基礎模型才說得通，而那又是另一個專案了。

Super VL 真正的用途是什麼

Read N​VIDIA's own framing and the target is narrow and specific: always-on enterprise agents that digest visual documents. The announcement's example is a web page — extract the text, read the charts, return findings — and the same claim is extended to PDFs, tables and structured documents. The argument for one model instead of a text model plus a separate vision pipeline is context continuity: the document, the data and the decision stay in the same window rather than being handed between systems.

那個論點在規格表上看來可信，因為上下文視窗達 100 萬個 token，而且視覺編碼器是聯合訓練而成，並非事後才硬加上去。至於它在 100 萬個 token 下能否經得起真實文件的考驗，正是 GraphWalks 與 MMLongBench-Doc 那些數字所要回答的問題，而這兩者都是廠商自行提報的數據。

N​VIDIA also lists a long adoption roster in the launch post — Scale AI building agents for Skydance and EY, AfterQuery for Hebbia's credit-agreement analysis, AI21 and Alice on cybersecurity, Mercor on coding agents, plus Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys among integrators. Those are partnerships named in a press release, which is a statement of intent rather than evidence of production traffic. Treat the roster as a signal about which verticals N​VIDIA is targeting, not as a reliability claim.

如何在不把生產路徑押在它上面的情況下試用它？

本週最誠實的立場是，Super VL 是一個有趣、文件記錄完善、但完全未經驗證的開放模型，而評估一個未經驗證的開放模型的正確方式，並不是在第一天就把正式環境流量導向它。

Locally, the NVFP4 build on a single H200 or a DGX Spark is the cheap test: pull the checkpoint, run your own document set through it, and compare against whatever vision model you use today. That costs you GPU time and nothing else, and it is the only way to get a number that is not N​VIDIA's. N​VIDIA's own documentation gives you one free head start: the model card and the technical report both specify temperature 1.0 and top_p 0.95 for text, and temperature 0.6 with top_p 0.95 for vision — a starting point you would otherwise spend a day rediscovering.

If you would rather test it behind an API, the availability line matters: Super VL is on Hugging Face, as N​VIDIA's own NIM microservice, and through several third-party platforms. N​VIDIA has published no price for any of those paths, so any per-token figure you see this week is set by the platform, not by the vendor. On the OrcaRouter side, N​emotron 3.5 Super VL is not on our catalogue — we do not serve it today, so there is nothing here to try it through yet, and we would rather say that than imply otherwise. Everything we do serve sits behind one key across 200-plus models at provider list price with 0% markup, and where a model is available from more than one provider, automatic failover across providers gives you a way to route to a new model on a fraction of traffic and let it fall back to the incumbent when it stumbles. That is the mechanism this kind of launch wants; it is not an offer to serve this particular checkpoint.

One more piece of the release worth a look regardless of whether you want the model: N​VIDIA published its post-training datasets alongside it — spans covering agentic tool use, finance, instruction following, long context, fairness and vision-language understanding — plus the RL environments in NeMo Gym. The dataset release is often more reusable than the checkpoint, and it is the part of this launch that costs N​VIDIA the most to give away.

看什麼

有四件事會改變這篇文章，大致依其影響程度排序。

第一個獨立基準測試，來自 Artificial Analysis 或任何其他人，要麼產出 5x 說法所暗示的 Super VL 那一列，要麼拒絕這麼做。在它存在之前，這次發布的核心效率宣稱，就只是新聞稿裡的一句話，後面掛著某個排行榜的名字，而但書應該和那個數字在同一口氣裡出現。

A price. N​VIDIA has told you the hardware, the runtimes and the licence, and withheld the one figure a buyer needs to compare this against a closed model. When NIM pricing appears, the comparison becomes arithmetic instead of positioning.

The Terminal-Bench 2.1 discrepancy — 66.70 on the model card against 60.67 in the technical report. Whichever is right, one of N​VIDIA's two primary documents is wrong, and a corrected re-run would settle several other rows at the same time.

And the walk from "available now" to "actually served." Super VL's benchmark rows compare it against Qwen3.5-122B-A10B, Qwen3.8-Flash-Next, DeepSeek-V4-Flash-0731, GLM-5.3-Flash and N​emotron 3 Ultra 550B A55B — all of which are mature enough to have measured throughput, real prices and production track records. A 12B-active model that ingests video, 1M-token documents and tool calls in one pass is a genuinely useful shape. Whether it earns a routing rule is a question about latency and price stability under load, and N​VIDIA has published the first of those only on its own hardware, under its own workload, in its own units.

基準測試會過時，價格會波動，但失敗紀錄不會：這一類模型背後有據可查的 AI 事故與代理失誤，全都收錄於the OrcaRouter AI Incident Archive，其中每一筆紀錄都附有自己的來源、嚴重程度、可信度評級與爭議標記。