Nemotron Sports Tennis vs North Micro Vision Instruct: A 31B Tennis Reader Against a 2.4B Document Specialist

Short answer first: NVIDIA NemotronLabs AI for Media - Sports Tennis and North Micro Vision Instruct are not substitutes, and nobody choosing between them is actually choosing between them. One is a 31B multimodal model that NVIDIA fine-tuned to answer questions about tennis points, needs roughly 80 GB of GPU memory, and reads English only. The other is a 2.4B vision-language model from Cohere that fits on a single workstation card, handles eleven languages, and was built above all to read documents — pages, charts, tables, OCR.

They sit in the same broad category and almost nowhere else. What follows is the honest version of the comparison: where the two genuinely diverge, what each vendor has and has not published, and the one situation where the choice is real.

What each model was built to do

North Micro Vision Instruct pairs a 2B language model — Cohere's North Micro LLM, following the Command A+ architecture — with a 400M-parameter vision encoder custom-trained from SigLIP 2 SO400M, for 2.4B total. Its vision path is native-resolution, preserving aspect ratios rather than resizing to a fixed grid, and a DeepStack projector injects patch embeddings from multiple encoder layers into the corresponding early language layers rather than prepending a single representation. Cohere's stated framing is a compact foundation for prototyping and task-specific fine-tuning, with document understanding as the flagship capability. The model card is unusually candid about the ceiling: North Micro Vision Instruct is explicitly not a reasoning model, has no tool calling, and was not trained with system prompts.

Sports Tennis is the opposite shape in every dimension. NVIDIA started from its own Nemotron 3 Nano Omni 30B-A3B-Reasoning checkpoint — a Mamba2-Transformer hybrid mixture of experts, 31B total with about 3B active per token — and fine-tuned it on a proprietary, manually labelled tennis corpus. The vision encoder (C-RADIOv4-H) and the speech encoder (Parakeet) were both frozen; only the language-model parameters moved. The result is a narrow model by design: it answers structured multiple-choice and open-ended questions over a full tennis point clip, spanning shot mechanics, court positioning, player movement, match facts, rules knowledge and audio cues. NVIDIA describes it as working best when an entire point — serve through the end of the rally — is passed as input.

The dimensions that actually separate them

• Parameters — North Micro Vision Instruct: 2.4B (2B language, 400M vision). Sports Tennis: 31B total, ~3B active per token.

• Inputs — North Micro Vision Instruct: interleaved text and images, native resolution, multi-image. Sports Tennis: video up to 2 minutes, audio up to an hour, images, text.

• Output — both return text. North Micro Vision Instruct additionally returns grounding bounding boxes on a normalised 0–1000 scale.

• Languages — North Micro Vision Instruct: eleven-plus, including English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese and Arabic. Sports Tennis: English only.

• Context — North Micro Vision Instruct: a 128K-token language backbone, but Cohere flags the validated multimodal range as up to 8K tokens, with longer multimodal contexts relying on extrapolation and unbenchmarked. Sports Tennis: up to 256k tokens.

• Footprint — North Micro Vision Instruct: about 4.97 GB at BF16, roughly 2.17 GB at 4-bit. Sports Tennis: about 62 GB at BF16, one 80 GB GPU required.

• License — North Micro Vision Instruct: Apache 2.0. Sports Tennis: OpenMDW-1.1, with the card stating commercial and non-commercial use.

Benchmarks: one model has them, the other has a protocol

This is where the two cards could not be more different in posture.

Cohere publishes numbers for North Micro Vision Instruct. DocVQA 0.921, ChartQA 0.808, OCRBench 0.792 — and MMMU 0.329. All vendor-reported, none independently reproduced, and Cohere itself flags that OCR results vary sharply by split. That last figure is worth pausing on: an MMMU score of 0.329 is low, and it is consistent with everything else the card says about the model's design. This is a reading specialist, not a general reasoning model, and the company that built it says so. That kind of disclosure is more useful than a flattering number.

NVIDIA publishes nothing. The Sports Tennis card describes its evaluation carefully — a test partition of 12 fully held-out matches, 2,689 point-level clips and 80,872 questions from matches with no training overlap, scored by automated multiple-choice accuracy and LLM-as-judge pass@9 on open-ended responses, with the seen-matches split explicitly excluded from reported testing — and then reports no result from any of it. The training side is equally detailed: 1,312,129 Q&A examples, 1,226,081 multiple-choice and 86,048 open-ended, drawn from 43,084 point-level clips across 239 matches, labelled along 38 annotation categories.

Even if NVIDIA had published scores, they would not be comparable. There is no benchmark on which a document-understanding model and a tennis-point model both appear. The overlap between these two evaluation stories is zero.

Deployment: where the practical difference lives

The 2.4B model is, by a wide margin, the easier thing to operate. Roughly 5 GB of BF16 weights means a single modern workstation GPU handles it, and the 4-bit build at around 2.17 GB goes smaller still. Independent ports exist for Apple's Core AI runtime, with a reported ~18.2 tokens/s at int8 on an iPhone 17 Pro and int4 quantisation failing outright. The friction is in the software: North Micro Vision Instruct requires Transformers 5.16.0 — installable from source until it reaches PyPI — and public vLLM support was still described as coming soon on the card. Fine-tuning is supported through Axolotl. There is no first-party hosted API; the practical path is self-hosting.

Sports Tennis is the harder thing to operate and there is no way around it. One 80 GB GPU at BF16, no quantized checkpoint published, English only, Linux only, on PyTorch/Transformers or NeMo or Megatron. NVIDIA tested across A100, H100, H200, B200, GB200 NVL72, RTX PRO 6000 SE, L40S, DGX Spark, Jetson Thor and RTX 5090 — a hardware list that says more about what NVIDIA wanted to validate than about what a normal team can provision. The card's default inference policy is bare-bones too: 2 frames per second up to 128 frames, 256 new tokens, greedy decoding.

Neither model is served on OrcaRouter. North Micro Vision Instruct has no first-party endpoint and isn't in our catalogue; Sports Tennis has no endpoint anywhere, since NVIDIA published weights and no hosted service. Both are self-hosting decisions.

Where a routing layer does help is the pipeline around them — whichever of these you run locally, the transcription, summarisation, retrieval and application models that surround it are hosted, and putting those behind one API key with automatic failover avoids standing up a separate credential set and contract for each. We pass provider list price through at 0% markup on the models we do carry, so the parts of the stack you don't self-host stay at vendor pricing.

When the choice is real

There is one scenario where picking between these two is a genuine decision, and it is worth being concrete about it because it is not obvious.

It is the case of a media or sports organisation that already processes documents and wants to add point-level video understanding. A broadcaster with an archive pipeline, a league with match reports and statistical PDFs, a rights holder with both text and footage — for them, North Micro Vision Instruct is plausibly already in the stack for OCR and chart extraction, and Sports Tennis is the new capability they would be adding. The question is not "which one" but "what does the second one cost me in infrastructure," and the answer is a datacenter GPU and an English-only constraint.

Outside that case, the models simply do not compete. If you need documents read in eleven languages on a workstation card, North Micro Vision Instruct is the answer and Sports Tennis is irrelevant to you. If you need to ask structured questions about tennis points with audio context, Sports Tennis is the only one of the two that does it, and the fact that it lacks published benchmarks is a risk you would be accepting rather than a reason to pick the other model.

Which to pick

Pick North Micro Vision Instruct if your work involves documents, charts, OCR, grounding or multilingual image understanding, and if you value a vendor that publishes its weak numbers alongside its strong ones. It is small, Apache 2.0, well documented, and honest about not being a reasoning model. Its MMMU of 0.329 is not a flaw you discover later; Cohere tells you up front.

Pick NVIDIA NemotronLabs AI for Media - Sports Tennis only if you specifically need point-level tennis reasoning with audio, you have an 80 GB GPU to spare, and you are comfortable being the first person to evaluate it. Nobody has published a score for this model, so the download is the experiment. That is not a reason to avoid it — a narrow specialist with a meticulously labelled 43,084-clip training set is exactly the kind of model that can beat a generalist at its own task — but it is a reason to treat the first deployment as a trial rather than a commitment.

The one thing you should not do is treat these as interchangeable because both say "vision-language model" on the card. A document reader and a tennis analyst were never the same product, and in this pair the difference in what they do is far larger than the difference in how well they do it.