Nemotron Sports Tennis vs Microsoft Mage-VL: Two Ways to Read a Sports Broadcast

Microsoft Mage-VL has a number you can point at: on the SoccerNet response-timing benchmark it posts a TimVal of 55.54 and a PR-AUC of 9.30, best on the precision-sensitive metrics despite never being trained on that dataset, and its authors demonstrate it sitting through a 2026 World Cup broadcast staying silent until 29.36 seconds, when a player breaks and the model fires live commentary. NVIDIA NemotronLabs AI for Media - Sports Tennis has no number at all. It is a 31B multimodal model fine-tuned on 43,084 tennis point clips that NVIDIA published in September 2026 with a full evaluation protocol on its card and not one result from it.

So this is not a head-to-head you can score. It is a genuinely useful comparison anyway, because the two models answer the same question — can a vision-language model follow live sport — with opposite architectural bets, and one of those bets is backed by evidence while the other is backed by a data description. That asymmetry is the article.

The fork: a stream, or a clip

The design difference matters more than the parameter counts, and it explains almost everything else about these two models.

Microsoft Mage-VL is built around continuous video. Its visual encoder, Mage-ViT, is trained from scratch and reads video the way a codec does: it splits a stream into anchor (I) frames, kept in full, and predicted (P) frames, where only the patches carrying real motion or new detail are retained, on a shared 16×16 patch grid. That cuts visual tokens by more than 75% and, per Microsoft, yields up to 3.5× wall-clock inference speedup over uniform frame sampling. On top of that sits a System 1 / System 2 split: a lightweight cognition gate scores each rolling window and stays silent on routine content, invoking the full model only when a response-worthy event completes. It is one model doing both jobs, not a pipeline.

NVIDIA's tennis model takes the other bet entirely. Its card is explicit that it "works best when a full tennis point clip is passed as input" — serve through the end of the rally, up to two minutes of mp4, with audio. The default inference policy is 2 frames per second up to 128 frames, 256 new tokens, greedy decoding, video plus audio. There is no gate, no streaming mode, no event trigger. It is a question-answering model over a bounded clip: you hand it a point, you ask what happened, it tells you.

Those are not two implementations of one idea. Streaming perception and clip-level reasoning are different products. A broadcaster wanting live commentary needs Mage-VL's shape. An analyst wanting to query an archive of 43,084 points needs Sports Tennis's shape. Neither model is a drop-in for the other's job.

Both are built on a hybrid backbone — with one important difference

• Parameters — Sports Tennis: 31B total, roughly 3B active per token, Mamba2-Transformer hybrid MoE. Mage-VL: 4B total, a 316M Mage-ViT paired with a Qwen3-4B-Instruct-2507 decoder.

• Encoders — Sports Tennis freezes both the C-RADIOv4-H vision encoder and the Parakeet speech encoder and fine-tunes only the language-model parameters. Mage-VL trains its entire visual stack from scratch on roughly 100M unlabeled images and videos, with the Qwen3 language model as the only pretrained component.

• Audio — Sports Tennis takes audio as a first-class input and lists audio-cue interpretation among its 38 annotation categories. Mage-VL's published pipeline is visual; the SoccerNet work is about response timing on frames.

• Modality out — both produce text only.

• Multiplier effect — because Mage-ViT was cheaper to pretrain than a web-scale vision tower, Microsoft reports training on videos 8× longer under the same budget. NVIDIA's efficiency claim is architectural instead: 3B active parameters per token out of 31B total.

Where the evidence sits

This is the part of the comparison that is not close, and it is worth stating plainly.

Microsoft Mage-VL is a published model with a technical report, a project page, a GitHub repository and a benchmark spread that covers image understanding, video understanding, temporal grounding, spatial reasoning and streaming. Against Qwen3-VL-4B — same 4B language backbone, only the vision encoder swapped — Mage-VL improves on every reported video and temporal-grounding benchmark, with the largest gains on localisation-heavy tasks: +22.5 on Timelens-QVHighlight, +17.1 on ActivityNet, +11.0 on VSI-Bench, +24.5 on VideoEval-Pro. It reports DocVQA 95.14, MMStar 67.32 and CrossPoint 80.00 on the image side, VideoMME 64.0 and LongVideoBench 61.3 on video, and a 64.00 overall score on OVO-Bench among streaming architectures. All of those are Microsoft-reported and none has been independently reproduced — but they exist, they are numerous, and they are internally consistent across an unusually wide set of tasks.

Sports Tennis reports nothing. Its card documents a test partition of 12 fully held-out matches with 2,689 point-level clips and 80,872 questions, describes scoring by automated multiple-choice accuracy and LLM-as-judge pass@9, notes that only the unseen-match split was used for reported testing — and then publishes no result. Its training corpus is real and specific: 1,312,129 Q&A examples, 1,226,081 multiple-choice and 86,048 open-ended, from 43,084 clips across 239 matches, labelled along 38 annotation categories from 2025 broadcast and court-capture footage. None of that is verifiable from outside, and the numbers that would make it verifiable are missing.

One caveat cuts the other way, and it is worth naming so this does not read as a clean win for Microsoft. SoccerNet is not tennis. Mage-VL's streaming credentials come from football broadcast data under a specific protocol, and its 2026 World Cup demonstration is a demonstration. Whether the codec-native gate transfers cleanly to tennis — where points are short, frequent and heavily structured — is untested. The difference is that Mage-VL's claim is at least falsifiable by a third party, and Sports Tennis's is not falsifiable by anyone without NVIDIA's dataset.

What it takes to run them

The gap here is roughly a factor of five in hardware, and it decides who can realistically try each model.

• Sports Tennis — one GPU with about 80 GB at BF16, weights around 62 GB. A100 80GB or H100 80GB at the floor, B200 or H200 recommended. No quantized checkpoint is published, so there is no cheap path down. English only. Runtimes are PyTorch/Transformers plus NeMo and Megatron.

• Mage-VL — roughly 10.6 GB at BF16, which puts a 16 GB GPU as the practical floor for image work and 24 GB or more for long video and streaming. Apache-2.0 for Mage-VL and MIT for Mage-ViT, though the family repository states the models are released for research purposes only. Note the sharp edges: trust_remote_code is required, there is no vLLM or SGLang serving path yet, and the codec pipeline needs FFmpeg or ffprobe — with the neural-codec path additionally needing DCVC-RT.

If you want to evaluate a sports-video model this month on a single workstation card, Mage-VL is the only one of the two you can actually load. That is a practical verdict even in the absence of a benchmark verdict.

Which one you would reach for

For anything live — commentary, event detection on a running feed, low-latency alerting on broadcast video — Microsoft Mage-VL is the defensible pick today: it was designed for exactly that, it has the streaming benchmark to argue for it, and it fits on hardware most teams already own. For archive-heavy, query-shaped work on tennis specifically — building a searchable index of points, running structured analysis across thousands of rallies, asking questions where the whole point clip is the unit of meaning — NVIDIA's model is the only one of the two that was built for it. Whether it does the job well is, as of September 2026, genuinely unknown.

There is a third option worth naming, because it is where most real pipelines end up. If you want to try the unproven specialist without betting a production path on it, keep it behind the same interface as the models you trust. OrcaRouter does not carry either of these — NVIDIA published weights with no endpoint, and Mage-VL has no hosted serving path — so both are self-hosting decisions. What a single key covering 200-plus hosted models buys you is the rest of the stack: the transcription, summarisation and application models around the sports-video component stay behind one endpoint, with automatic failover if a provider degrades, and the two specialist models you run yourself are the only moving parts you own.

The verdict

Microsoft Mage-VL and NVIDIA NemotronLabs AI for Media - Sports Tennis are not rivals in the sense a versus page usually means. Mage-VL is a published, benchmarked, 4B streaming model that runs on a workstation and has a real demonstration of event-gated sports commentary. Sports Tennis is an unannounced, unbenchmarked, 31B clip-level specialist that needs a datacenter GPU and carries no published evidence it works.

Judge on the evidence and Mage-VL wins by forfeit. Judge on scope and they do not overlap. But there is a reason to keep watching NVIDIA's model: a frozen-encoder fine-tune over 43,084 correctly-annotated tennis points is exactly the kind of narrow, high-quality vertical training set that generalist models do not have, and if NVIDIA ever publishes the held-out results, the specialist-versus-generalist question in sports video gets its first real answer. Until then, Mage-VL is the model with receipts and Sports Tennis is the model with a promise.