Nemotron Parse 2.0 vs olmOCR: Two Jobs, Both Called Parsing

On August 3, 2026, NVIDIA posted a new document-parsing model to Hugging Face and told nobody. NVIDIA-Nemotron-Parse-2.0 is a 0.9B vision-encoder-decoder whose model card claims a multilingual and chart-aware jump over its February 2026 predecessor, and it landed in the same corner of the ecosystem as olmOCR — more precisely olmOCR-2-7B-1025, the 7B linearizer from the Allen Institute for AI that has quietly become the default way to turn PDFs into LLM training data. Calling this a head-to-head at all is the trap: both models get described as "document parsing," but they return different artifacts, feed different pipelines, and their benchmark numbers never actually face each other. The real question is which of two jobs you are hiring a parser to do.

The two artifacts

Nemotron Parse 2.0 is a layout-semantics engine. Give it a page image and a task prompt and it returns the text plus a semantic layer on top: a layout class for every region (title, section, caption, table, chart, header, footer, footnote, bibliography entry), a bounding box for each, and the reading order between them — with markdown as one optional serialization on top. olmOCR 2 is a linearizer. It takes the same page and returns clean, linearized markdown with a short YAML frontmatter recording page-level metadata such as primary language, rotation correction, and whether the page is a table or a diagram. No boxes, no classes, no geometry: one pass, text in document order.

The artifact determines the job. If your pipeline needs to cite a fact back to a region of the page, highlight a table on a scan, or extract layout semantics for document intelligence, you need geometry — that is Nemotron Parse 2.0's output space. If you are building training data or feeding a text LLM that only consumes tokens, geometry is noise and linearized markdown is exactly right — that is olmOCR's whole design. You can bolt layout detection onto olmOCR's output with a separate detector, and you can discard Nemotron Parse 2.0's boxes and keep just the markdown, but each model is built to make one of those jobs native.

The quiet ship: what the repo shows, what is not confirmed

NVIDIA-Nemotron-Parse-2.0 appeared on Hugging Face on August 3, 2026 with no announcement, no blog post, and no NIM packaging. What is knowable is the repo. It is a 0.9B vision-encoder-decoder: a ViT-H image encoder built on NVIDIA's C-RADIO backbone, a lightweight 1D-convolution adapter, and a ten-layer mBART-style decoder. The tokenizer grew by roughly 20,000 entries to about 72,000 total, explicitly for more efficient multilingual support, and the card lists chart-aware parsing as a headline feature via a new class_Chart token, plus a family of table serializations (LaTeX, HTML, markdown, JSON, hierarchical JSON, CSV). Two optional logits processors ship with it: one that stops repetitive structured output, and one that forces a table structure on a table crop. Recommended input resolution spans about 1024×1280 up to 1664×2048, generation runs to a 9,000-token ceiling, and the model card lists Transformers, vLLM, SGLang, and Docker Model Runner as runtimes on Ampere, Hopper, Blackwell, and Turing hardware.

The claims, however, are all NVIDIA's own, and none have been independently reproduced. The card reports ParseBench overall at 0.6391 (up from v1.2's 0.5782), MOSCAR multilingual OCR at 0.9102 BoC-F1 (up from 0.4410), IndicVisionBench at 0.7203 ANLS-character (up from 0.0612), and an OmniDocBench handwriting subset improving from 0.9739 to 0.3395 on text-edit distance. Those are the largest jumps, and they are also the least verified: ParseBench is NVIDIA's own suite, and no third party has run Parse 2.0 against an independent corpus yet. What is not confirmed goes beyond scores — there is no hosted inference (the card states the model is not deployed by any inference provider), no pricing, and no announced timeline for either.

olmOCR 2: the incumbent everyone already measures by

olmOCR is the model that invented the benchmark this category now argues over. olmOCR-2-7B-1025, released October 22, 2025, is a 7B vision-language model fine-tuned from Qwen2.5-VL-7B-Instruct on the 270,000-page olmOCR-mix-1025 corpus, then refined with GRPO reinforcement learning against automated "unit test" rewards — deterministic checks that a table kept its structure, that a formula transcribed exactly, that reading order held. That unit-test framing became olmOCR-Bench, roughly 1,400 PDFs and more than 7,000 checks, and it has become the de facto industry benchmark: Marker, MinerU, and a dozen commercial OCR models now publish on it. olmOCR 2 itself scores 82.4 overall there, about four points above the previous release and above the Marker (76.1) and MinerU (75.8) figures AI2 cites on the same page.

olmOCR is also the mature option in this pairing. It is Apache-2.0, its weights have been downloaded roughly 218,000 times in the last month, and the olmOCR toolkit handles rendering, rotation, and retries at scale on a vLLM backend — AI2's batch estimate puts the pipeline around $176–190 per million pages on rented GPUs, and the FP8 build runs about 3,400 output tokens per second on a single H100, fast enough that the release post quotes "10,000 pages for less than $2." The tradeoff is language: olmOCR is explicitly built and benchmarked for English-language digitized print, and its model card tags it English. Nemotron Parse 2.0's whole pitch is the opposite — its claimed gains are concentrated in CJK and Indic scripts.

Six rows where the tradeoff shows

Both models are open-weight, both run on Transformers, vLLM, or SGLang, and both are self-hosted today. On the dimensions that matter for choosing between them:

• Size — Nemotron Parse 2.0 is 0.9B; olmOCR 2 is 7B. Roughly eight times the parameters, roughly eight times the VRAM floor.

• Output — Nemotron Parse 2.0 returns text, layout classes, bounding boxes, reading order, and markdown; olmOCR 2 returns linearized markdown with a YAML metadata block.

• Multilingual — Nemotron Parse 2.0 claims strong CJK and Indic support (MOSCAR 0.9102, IndicVisionBench 0.7203, vendor-reported); olmOCR 2 is English-first.

• Flagship score — Nemotron Parse 2.0 reports ParseBench 0.6391 (NVIDIA's own, unaudited); olmOCR 2 scores 82.4 on olmOCR-Bench (AI2's own, ten months of community reuse).

• License — Nemotron Parse 2.0 ships under the NVIDIA Open Model License; olmOCR 2 is Apache-2.0.

• Shipped — Nemotron Parse 2.0 arrived August 3, 2026, unannounced; olmOCR 2 shipped October 22, 2025, with a launch post.

Three places the decision actually bites

Scale and latency. A 0.9B decoder is dramatically cheaper to serve than a 7B VLM: a smaller GPU, less VRAM, more pages per second on the same hardware, and a lower cost per page once you are paying for GPU time. If you already run GPU infrastructure and your corpus is large, that is a real monthly difference. olmOCR's own numbers show how fast a 7B model can be made with FP8 quantization — but it still needs an H100-class GPU to hit them; Nemotron Parse 2.0's hardware floor is an Ampere-era card.

Multilingual. This is the most asymmetric row. Nemotron Parse 2.0 grew its tokenizer by about 20,000 entries specifically for multilingual OCR, and its card's claimed gains are concentrated in Indic scripts and CJK. olmOCR 2 makes no such claim — its language tag is English. If your corpus is Hindi, Tamil, Bengali, Japanese, or mixed-script, Nemotron Parse 2.0's numbers are the ones worth testing, precisely because they are vendor-reported and fresh.

The serving reality. olmOCR 2 is ten months old and its toolchain is boring in the good way. Nemotron Parse 2.0 is five days old and its serving story is visibly young: vLLM needs a build with Nemotron Parse remote-code support, the tied output head trips some vLLM 0.20 builds (the repo ships a runtime patch), and the tokenizer.json carries serialized padding to 9,000 tokens — one serving stack hit a six-token prompt expanding to 54,000 token IDs before patching. None of that is fatal, but it is exactly the kind of friction you budget for with a new model.

Pick one for your pipeline

Choose olmOCR 2 if you are building LLM training data or a high-volume text-extraction pipeline over English documents. You get a mature toolkit, an Apache-2.0 license, the benchmark the industry now measures by, and a well-trodden batch path. Its accepted failure mode is language coverage.

Choose Nemotron Parse 2.0 if your pipeline needs geometry — layout classes, bounding boxes, and reading order for grounded retrieval or document intelligence — or if your corpus is heavy on Indic, CJK, or handwritten pages, where its claimed gains sit. The 0.9B size makes a weekend test cheap even if you are unsure. Its accepted failure mode is that every number on the card is NVIDIA's own and could move under independent evaluation.

If your input is mostly English born-digital PDFs and all you need is clean markdown, olmOCR 2 is the lower-risk default. If you are already self-hosting and the multilingual or layout-grounded use case is real, Nemotron Parse 2.0 is the more interesting bet — test it on your own pages before committing.

Keep the parser choice from becoming a lock-in

A document pipeline has a parser stage and then an LLM stage, and the parser is usually the cheapest stage once volume is real — the LLM that turns parsed markdown into summaries, structured records, or answers is where cost scales. That is the stage a routing layer changes. OrcaRouter puts 200+ models behind one API at the provider's list price with no markup, so a vendor price cut is live the same day, and automatic failover keeps one degraded provider from stalling a batch job. Neither parser in this comparison is on our catalog — both model cards state they are not deployed by any inference provider, so this is a self-host decision — but the reason to keep the parser decoupled from your model stack is exactly what routing buys on the downstream side: swap Nemotron Parse 2.0 for olmOCR 2, or back, without touching a single integration, because the LLM layer stays behind the same one-key API.

FAQ

Which model wins on benchmarks?

Neither — the numbers do not face each other. Nemotron Parse 2.0 reports ParseBench, MOSCAR, IndicVisionBench, and an OmniDocBench handwriting subset, all NVIDIA's own and none independently reproduced. olmOCR 2 reports olmOCR-Bench, AI2's own benchmark that the rest of the industry now publishes on. No third party has run both models on the same corpus, so any direct "winner" claim is extrapolation. Directionally: olmOCR's figures have survived ten months of community use; Nemotron Parse 2.0's are fresh and could move.

Is Nemotron Parse 2.0 really better at non-English documents?

That is the claim — a roughly 20,000-token vocabulary expansion, plus MOSCAR at 0.9102 and IndicVisionBench at 0.7203, both vendor-reported and both up sharply from v1.2. olmOCR 2 makes no multilingual claim and is tagged English. But until an independent run lands, treat Nemotron Parse 2.0's multilingual gains as promising-but-unverified and test on your own Indic or CJK pages before relying on them.

Do I need a bigger GPU for one of them?

Yes, and it tracks the size difference. Nemotron Parse 2.0's 0.9B fits on a modest Ampere-era card and is the cheaper per-page option on hardware you already own. olmOCR 2's 7B VLM wants a substantially larger GPU; its FP8 build reaches about 3,400 output tokens per second on an H100, per AI2.

Which is better for RAG preprocessing?

It depends on what your retriever needs. If you ground answers back to page regions, need layout classes, or want to index tables and charts as their own units, Nemotron Parse 2.0's bounding boxes and classes give you that natively. If your RAG stack just consumes clean text and your corpus is English, olmOCR 2's linearized markdown is proven, simpler, and backed by a mature toolkit.

Bottom line

NVIDIA shipped a real parser this week and told no one; AI2 shipped one ten months ago and gave the category its benchmark. Nemotron Parse 2.0 is the better fit when you need layout semantics, multilingual coverage, or handwriting extraction on a small budget — and the areas where its numbers are least verified are exactly the areas it claims to win. olmOCR 2 is the better fit when you need linearized text at scale over English documents and want a toolchain that has been stable for ten months. Neither is deployed anywhere, so this is a self-host decision either way; the cheapest way to make it is to run both on your own hardest pages and watch where each breaks.