Best Summarization Models for Apple Silicon Macs in 2026: BART Large CNN and PEGASUS XSum

Quick Answer

As of September 2026, BART Large CNN (facebook) is the top pick under “newest release first, then downloads,” winning the download tiebreak among equally dated candidates; too few recent releases qualify, so the newest available are listed, with no public benchmark yet for any candidate [1][3][5][7]. In order, the ranking is BART Large CNN (facebook), PEGASUS XSum (google), BigBird PEGASUS Large arXiv (google), and BART Large XSum (facebook) [1][3].

Key Takeaways

• Ranking note: Scope admits only summarization-tagged models from labs with a published paper or leaderboard record. [2][4][6] Too few recent releases qualify, so the newest available candidates are included; all were first published on Hugging Face on 2022-03-02. [1][3][5][7] No public benchmark scores any candidate: ordering is newest release first, then downloads. The established picks—BART Large CNN, PEGASUS XSum and BigBird PEGASUS Large arXiv—follow the same rule; their exemption does not determine first place. [1][3][5]
• facebook’s BART Large CNN (facebook/bart-large-cnn) ranks first because the publication dates tie and its 1,358,863 downloads win the popularity tiebreak; that position does not establish better summarization quality. [1][3][5][7] Context: 1K tokens; license: MIT; no public benchmark yet. [1]
• google’s PEGASUS XSum (google/pegasus-xsum) ranks second with 198,507 downloads and a 512-token context; check input length before choosing it for document summaries. No public benchmark yet. [3]
• google’s BigBird PEGASUS Large arXiv (google/bigbird-pegasus-large-arxiv) ranks third with 9,146 downloads; its 4K-token context provides more input room than the other listed candidates. [1][3][5][7] License: Apache-2.0; no public benchmark yet. [5]
• facebook’s BART Large XSum (facebook/bart-large-xsum) ranks fourth with 6,619 downloads, a 1K-token context and an MIT license; no public benchmark yet. [7]
• BART Large CNN has 406M parameters [1], implying 203 MB for 4-bit weights, estimated (params x 0.5 bytes). [1] Weight storage alone does not establish runtime memory requirements or performance on Apple Silicon Macs.

How do these summarization models compare on memory, context, licenses and published benchmarks?

Which summarization models should you consider for an Apple Silicon Mac?

1. BART Large CNN

BART Large CNN by facebook ranks first because its Hugging Face publication date matches the other candidates and its download count breaks the tie.[1][3][5][7] The model has 406M parameters, a 1K-token context window, and an MIT license.[1] Its practical use here is summarizing material that fits within that input limit; the ranking does not establish superior summary quality. Benchmark status: no public benchmark yet.

For an Apple Silicon Mac, quantized weight storage would be approximately 203 MB at 4-bit precision, estimated (params x 0.5 bytes) from the documented 406M parameters.[1] Published F32 weights occupy 1.6 GB.[1] Weight storage alone does not establish total memory needed to run locally: runtime allocations also need room. A specific Mac or unified-memory configuration cannot be recommended from those figures alone. Treat the quantized figure as a planning estimate, without assuming an available, compatible quantized package.

Ranking note: the ordering is newest release first, then downloads. The shortlist uses the newest available entries because too few recent releases qualify; every listed candidate was first published on Hugging Face on March 2, 2022.[1][3][5][7] The ranking admits only summarization models from labs with a published paper or leaderboard record, using Hugging Face’s summarization pipeline tag. BART Large CNN’s 1,358,863 downloads over the preceding 30 days, measured September 25, 2026, serve only as the publication-date tiebreak.[1]

2. PEGASUS XSum

PEGASUS XSum by google occupies second place as an established pick, with 198,507 downloads over the preceding 30 days as of September 25, 2026.[3] The ordering is newest release first, then downloads: all candidates share a March 2, 2022 Hugging Face publication date, so downloads resolve the tie.[1][3][5][7] PEGASUS XSum has no public benchmark yet for this comparison; its position does not establish better summarization quality. Its publication date also places it outside the requested recent-release window.[3]

The context limit is 512 tokens, making short-input summarization the practical use case to evaluate.[3] Choose inputs that fit within that budget when assessing whether its summaries preserve the information you need. Longer documents require a separate input-handling strategy; the context limit alone does not establish how well summaries will retain details across a complete document. PEGASUS is documented in google’s paper, “PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization.”[4]

For local deployment on an Apple Silicon Mac, hardware requirements remain unverified here. Parameter count, quantized weight size, license, and a compatible local runtime are not specified in the available model details.[3] A defensible memory estimate therefore cannot be given. Treat PEGASUS XSum as a candidate for evaluation, with its short context as the concrete constraint, and verify runtime compatibility and licensing before committing to deployment.

3. BigBird PEGASUS Large arXiv

BigBird PEGASUS Large arXiv from google ranks third under the ordering rule “newest release first, then downloads”: the listed candidates share a Hugging Face publication date, and its download count places it here.[1][3][5][7] Its first publication was March 2, 2022, with 9,146 downloads over the preceding 30 days as of September 25, 2026.[5] The shortlist uses the newest available options because too few qualify within the requested release window. Eligibility is limited to summarization-tagged models from labs with a published paper or leaderboard record.

BigBird PEGASUS Large arXiv has a 4K-token context window and an Apache-2.0 license.[5] Consider it for academic-document summarization when the passage you want to summarize fits that window. The context allowance gives it more room for input than the listed BART and PEGASUS XSum alternatives, although context capacity alone does not establish summary quality.[1][3][5][7] Its associated paper is Big Bird: Transformers for Longer Sequences.[6]

Apple Silicon hardware requirements remain unverified here: no parameter count, quantized weight size, or Mac runtime measurement establishes a defensible memory recommendation. A specific RAM configuration therefore cannot be promised to fit or run well. Check compatibility and memory use in your intended local runtime before committing to it. The central caveat is no public benchmark yet for this ranking; its position reflects publication timing and downloads, rather than demonstrated summarization performance.

4. BART Large XSum

BART Large XSum by facebook ranks fourth under the ordering rule: newest release first, then downloads.[1][3][5][7] All listed candidates share a Hugging Face publication date of 2022-03-02; its 6,619 downloads over the last 30 days place it last on that tiebreak.[1][3][5][7] BART Large XSum has no public benchmark yet, so its position does not establish summarization quality. The shortlist uses older available releases because qualifying recent releases are too few. Eligibility is limited to summarization-tagged models from labs with a published paper or leaderboard record.

BART Large XSum provides a 1K-token context window and an MIT license.[7] Short-input summarization is the practical use case suggested by that context limit: check the tokenized input before selecting it for a workflow. Documents exceeding the window need an explicit handling strategy, such as splitting the input. The context ceiling is the concrete caveat for document summarization; the ranking provides no benchmark basis for promising better summaries than another candidate.

For local use on an Apple Silicon Mac, a defensible hardware recommendation remains unavailable. A verified parameter count, quantized weight size and runtime memory measurement are needed before specifying a RAM requirement. No weight-memory estimate is justified here, and local runtime compatibility remains unverified. Confirm that the intended runtime supports the checkpoint, then measure memory consumption with representative inputs before committing to deployment.

How can you estimate model weight memory from parameter counts?

Estimate model weight memory by multiplying the parameter count by the storage per parameter; for 4-bit weights, label the calculation “estimated (params x 0.5 bytes)” alongside the cited parameter count.[1]

Facebook’s BART Large CNN (facebook/bart-large-cnn) has 406 million parameters.[1] Its 4-bit weight memory is approximately 203 MB, estimated (params x 0.5 bytes) from that parameter count.[1] Treat that result as a planning estimate for the weights, rather than a measured memory requirement for running summarization on a Mac.

Keep the estimate separate from the published checkpoint size. The model card lists F32 weights at 1.6 GB.[1] That figure describes the published weights; the calculated 4-bit figure does not establish that a compatible quantized checkpoint is available or that a particular local runtime supports it.[1]

For a practical comparison, record the parameter count, assumed precision and estimated weight memory together. Leave weight memory unspecified when a parameter count is unavailable. Avoid turning the weight estimate into a claim about required Mac RAM: establishing whether a model fits and runs well requires a runtime memory measurement, which the calculation alone cannot provide.

How do context limits affect your choice of summarization model?

Context limits determine whether you can submit a document in one pass or need to divide it before summarization, so choose a model whose input limit suits your documents.

Google’s google/pegasus-xsum has a context limit of 512 tokens.[3] Facebook’s facebook/bart-large-cnn and Facebook’s facebook/bart-large-xsum each support 1K tokens.[1][7] Google’s google/bigbird-pegasus-large-arxiv supports 4K tokens, giving it the largest context window among these candidates.[5][1][3][7] For documents that exceed the BART models’ limits but remain within BigBird-Pegasus’s limit, that additional capacity provides a practical reason to consider BigBird-Pegasus.[1][5][7]

Check your document’s token count with the intended model’s tokenizer before choosing a workflow. For input beyond the model’s limit, plan to split the document at meaningful boundaries, summarize the sections, and combine their summaries. Keep headings and relevant surrounding material with each section, then review the combined result for missing connections. Avoid silently truncating a document when the summary needs to cover its entirety.

On an Apple Silicon Mac, evaluate context capacity separately from memory requirements. Treat the published context limit as an input constraint, not a guarantee of runtime performance or summary quality. Start with representative documents and check whether your chosen workflow preserves the information you need.

Why does this ranking include older models, and how are they ordered without public benchmark scores?

Older models are included because this summarization family has too few recent releases; the newest available candidates are listed, with the ordering rule “newest release first, then downloads.”[1][3][5][7] The ranking admits only summarization models from labs with a published paper or leaderboard record, using Hugging Face’s summarization pipeline tag.[1][2][3][4][5][6][7]

All listed candidates were first published on Hugging Face on March 2, 2022, so that date leaves them tied.[1][3][5][7] Downloads therefore break the tie: Facebook’s facebook/bart-large-cnn has 1,358,863; Google’s google/pegasus-xsum has 198,507; Google’s google/bigbird-pegasus-large-arxiv has 9,146; and Facebook’s facebook/bart-large-xsum has 6,619.[1][3][5][7] Those counts cover the preceding 30 days as of September 25, 2026.[1][3][5][7]

Facebook’s facebook/bart-large-cnn takes first place because its publication date ties with the other candidates and its download count wins the tiebreak, not because a benchmark establishes superior summarization quality.[1][3][5][7] The established picks—the family’s three most-downloaded models—are facebook/bart-large-cnn, google/pegasus-xsum, and google/bigbird-pegasus-large-arxiv; their recency exemption does not itself award first place.[1][3][5]

Every entry carries the status “no public benchmark yet”: no public leaderboard scores these candidates against one another. The order consequently provides no measured comparison of summary quality or Apple Silicon performance. Model size does not determine placement, and popularity serves only as the stated tiebreak.

Frequently Asked Questions

Which summarization model should I try first on an Apple Silicon Mac?

facebook/bart-large-cnn from Facebook heads this list because the candidates share a publication date and its download count breaks the tie.[1][3][5][7] The remaining order is google/pegasus-xsum from Google, google/bigbird-pegasus-large-arxiv from Google, then facebook/bart-large-xsum from Facebook.[3][5][7] Treat that ordering as a starting point for evaluation: download counts do not establish summarization quality, Apple Silicon compatibility, or performance on your Mac.

How current is this ranking, and how were the models ordered?

This ranking admits only summarization models from labs with a published paper or leaderboard record, using Hugging Face’s summarization pipeline tag.[2][4][6] The recent-release pool is too thin, so the newest available eligible entries are listed; all were first published on Hugging Face on March 2, 2022.[1][3][5][7] With no public benchmark scoring these candidates, the ordering is newest release first, then downloads.[1][3][5][7] No separate established-pick exemption changes that order.

How much memory would quantized BART Large CNN need?

For facebook/bart-large-cnn, 406M parameters imply approximately 203 MB of weight storage at 4-bit, estimated (params x 0.5 bytes).[1] The listed F32 weights occupy 1.6 GB.[1] The calculated weight size does not establish total runtime memory, availability of a compatible quantized implementation, or whether a particular Mac configuration can run the model. Use it as a weight-storage estimate, not a complete memory budget.

Which listed model has room for longer documents?

google/bigbird-pegasus-large-arxiv has a listed context length of 4K tokens.[5] Both facebook/bart-large-cnn and facebook/bart-large-xsum list 1K tokens, while google/pegasus-xsum lists 512 tokens.[1][7][3] Those limits make BigBird PEGASUS the candidate to evaluate when input length drives your choice. Context capacity alone does not establish summary accuracy, runtime memory requirements, or how quickly the model processes a document on Apple Silicon.

Are there published benchmark scores proving which model summarizes better?

The benchmark status for each listed candidate is “no public benchmark yet.”[1][3][5][7] Consequently, this comparison cannot identify a measured summarization-quality winner or an Apple Silicon speed winner. The download-based tie-break should not be read as either result. For a practical evaluation, compare summaries of your own documents for factual accuracy, important omissions, and adherence to your requested format.

What licenses do these summarization models use?

facebook/bart-large-cnn and facebook/bart-large-xsum list the MIT license.[1][7] google/bigbird-pegasus-large-arxiv lists Apache-2.0.[5] A license is not specified for google/pegasus-xsum in this comparison, so its licensing status remains unresolved here.[3] Keep that distinction visible when selecting a model for a product: the listed license information supports evaluating those named checkpoints, but does not establish the terms for every associated dependency or distribution.

Sources

1. facebook/bart-large-cnn model card (Hugging Face) — 2026-09-25
2. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension — 2019-10-29
3. google/pegasus-xsum model card (Hugging Face) — 2026-09-25
4. PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization — 2019-12-18
5. google/bigbird-pegasus-large-arxiv model card (Hugging Face) — 2026-09-25
6. Big Bird: Transformers for Longer Sequences — 2020-07-28
7. facebook/bart-large-xsum model card (Hugging Face) — 2026-09-25