Best Summarization Models for CPU-Only PCs in 2026: BART Large CNN and PEGASUS XSum

Quick Answer

BART Large CNN (facebook) is the top pick as of September 2026 under “newest release first, then downloads”: the available candidates are older releases with a shared publication date and no public benchmark yet, so downloads break the tie [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 scope: only summarization-tagged models from labs with a published paper or leaderboard record qualify. The recent-release pool is too small, so the newest available entries are listed; all have a Hugging Face publication date of March 2, 2022.[1][2][3][4][5][6][7]
• Ranking note: no public benchmark scores any candidate, so the ordering is “newest release first, then downloads.” The established picks are Facebook’s BART Large CNN, Google’s PEGASUS XSum and Google’s BigBird-Pegasus Large arXiv; their recency exemption does not itself grant first place.[1][3][5][7]
• 1. Facebook’s BART Large CNN ranks first because publication dates tie and its download count wins the specified tiebreak.[1][3][5][7] Its 406M parameters imply 203 MB of 4-bit weights, estimated (params x 0.5 bytes); context is 1K tokens, the license is MIT, and there is no public benchmark yet.[1]
• **2. Google’s PEGASUS XSum (google/pegasus-xsum) has a 512-token context and no public benchmark yet [3]. Parameter count, weight-memory requirements and license are unspecified here, leaving those deployment checks unresolved.[3]
• **3. Google’s BigBird-Pegasus Large arXiv provides a 4K-token context under the Apache-2.0 license [5]. Its context exceeds the other listed candidates, but parameter count and weight-memory requirements are unspecified here; no public benchmark yet.[1][3][5][7]
• **4. Facebook’s BART Large XSum offers a 1K-token context under the MIT license, with no public benchmark yet [7]. Its parameter count and weight-memory requirements are unspecified here, so a CPU RAM requirement cannot be established.[7]

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

Which summarization models should you consider for a CPU-only PC?

1. BART Large CNN

BART Large CNN by facebook ranks first because its publication date ties the other candidates and its download count breaks the tie, rather than demonstrating superior summarization quality.[1][3][5][7] Ranking note: newest release first, then downloads. The shortlist has too few releases within the past twelve months, so the newest available are listed; BART is an established pick, published on Hugging Face on March 2, 2022.[1] Scope: this ranking admits only summarization models from labs with a published paper or leaderboard record, using Hugging Face’s summarization pipeline tag.

BART Large CNN has 406 million parameters, a context length of 1K tokens, F32 weights totaling 1.6 GB, and an MIT license.[1] Quantized weight storage would be approximately 203 MB at 4-bit precision—estimated (params x 0.5 bytes) from the cited 406 million parameters.[1] That estimate covers weights alone. A CPU-only PC also needs memory for execution, so neither weight figure establishes a minimum system RAM requirement. A working local configuration still requires verification of runtime and quantization support.

Use BART Large CNN for summarization jobs whose input fits its 1K-token context; longer documents require a separate input-handling strategy.[1] The key caveat is validation: no public benchmark yet. CPU throughput and total runtime memory requirements are also unverified, so its ranking should guide evaluation rather than promise a particular speed or hardware fit.

2. PEGASUS XSum

PEGASUS XSum from google occupies the second position under the ordering rule: newest release first, then downloads.[1][3][5][7] The listed candidates share a Hugging Face publication date, so downloads determine their order; popularity does not establish summarization quality.[1][3][5][7] PEGASUS XSum has no public benchmark yet. The ranking admits only summarization-tagged models from labs with a published paper or leaderboard record. Too few recent releases qualify, so the newest available candidates are listed.

The google/pegasus-xsum model card documents a context length of 512 tokens, a Hugging Face publication date of March 2, 2022, and 198,507 downloads over the reporting window ending September 25, 2026.[3] Its methodological foundation is google’s “PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization.”[4] Consider it for abstractive summarization of short inputs that stay within the documented context limit.[3][4] Longer documents require an input-handling decision before model selection.

For local CPU use, a concrete RAM recommendation remains unverified: parameter count, quantized weight size, runtime compatibility and CPU throughput are not established here. A defensible quantized-memory estimate therefore cannot be supplied. Confirm those requirements and the license before adopting the checkpoint. The practical caveat is its limited input window: the documented context supports evaluating short-input workflows, but does not establish suitability for summarizing an entire long document in a single pass.[3]

3. BigBird-Pegasus Large arXiv

BigBird-Pegasus Large arXiv by google ranks third as an established pick, with a Hugging Face publication date of March 2, 2022, and 9,146 downloads over the last 30 days as of September 25, 2026.[5] The ordering is newest release first, then downloads: its publication date matches the preceding candidates, while its download count is lower.[1][3][5] Download popularity determines placement here; it does not demonstrate better summarization quality.

The model provides a 4K-token context window and an Apache-2.0 license.[5] Consider it for document summarization when your input exceeds the shorter context windows of the other listed candidates but remains within its supported context.[1][3][5][7] Hardware sizing needs a separate check: a parameter count and quantized checkpoint size are not specified here, so a defensible estimate for quantized weight memory or total system RAM cannot be given. Validate CPU runtime support, memory consumption and latency before committing to a local deployment.

The benchmark status is no public benchmark yet, so this placement does not establish summary quality or CPU speed. The ranking admits only summarization models from labs with a published paper or leaderboard record, using Hugging Face’s summarization pipeline tag. The shortlist uses the newest available candidates because too few recent releases qualify; BigBird-Pegasus belongs to the established picks retained beyond that recency window.[1][3][5][7] Its associated paper is Big Bird: Transformers for Longer Sequences.[6]

4. BART Large XSum

BART Large XSum by facebook ranks fourth in this selection, with 6,619 downloads over the last 30 days as of September 25, 2026.[7] The ordering is newest release first, then downloads: no public benchmark scores any candidate. All candidates share a March 2, 2022 Hugging Face publication date, so downloads determine their positions.[1][3][5][7] The family has too few recent releases; the newest available candidates are listed instead. BART Large XSum has no public benchmark yet.

The model has a 1K-token context window and an MIT license.[7] BART’s architecture is described in facebook’s paper, “BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension.”[2] This ranking admits only summarization models from labs with a published paper or leaderboard record, using the Hugging Face summarization pipeline tag.[1][2][3][4][5][6][7] Consider BART Large XSum for short-input summarization where the MIT license suits your project.[7]

For local CPU use, RAM requirements, quantized weight memory and processing speed remain unverified; a specific hardware configuration cannot be recommended confidently. Check runtime compatibility and measure memory use on your intended PC before committing to deployment. The practical caveat is input length: documents exceeding the 1K-token context window require an input-handling strategy, such as splitting the text before summarization.[7] Evaluate the resulting summaries on representative documents before relying on that workflow.

Why does the ranking include older summarization models?

The ranking includes older summarization models because too few recent releases meet the eligibility rules, so the newest available eligible checkpoints remain in the selection.[1][3][5][7] The listed Hugging Face publication dates are all March 2, 2022, placing every candidate outside the intended twelve-month release window.[1][3][5][7] Their inclusion is an explicit age exception, not evidence of a recent model generation.

The ranking admits only summarization models from labs with a published paper or leaderboard record, using Hugging Face’s summarization pipeline tag. Facebook’s BART models, Google’s PEGASUS model and Google’s BigBird-Pegasus model have published papers documenting their underlying model families.[2][4][6]

The ordering basis is exactly: newest release first, then downloads. No public benchmark scores any candidate against the others here, so each carries “no public benchmark yet.” Facebook’s facebook/bart-large-cnn takes the first position because the publication dates tie and its download count is highest among these candidates.[1][3][5][7] Google’s google/pegasus-xsum, Google’s google/bigbird-pegasus-large-arxiv and Facebook’s facebook/bart-large-xsum follow in that order.[3][5][7]

The established picks are facebook/bart-large-cnn, google/pegasus-xsum and google/bigbird-pegasus-large-arxiv, the selection’s three most-downloaded models.[1][3][5][7] Their exemption permits inclusion; it does not award the first position. Downloads resolve the release-date tie, without establishing summary quality or CPU speed. Readers should treat the order as a transparent selection rule, not a measured CPU performance ranking.

How can you estimate memory requirements from parameter counts?

Estimate weight memory by multiplying the parameter count by the storage per parameter; for 4-bit weights, use an estimate of 0.5 bytes per parameter.[1] Treat the result as a weight-storage estimate, not a complete system RAM requirement.

Facebook’s facebook/bart-large-cnn has 406 million parameters.[1] Its calculated weight storage is approximately 203 MB at 4-bit precision, estimated (params x 0.5 bytes) from that parameter count.[1] The published F32 weights occupy 1.6 GB.[1] Keep the published file size and the calculated quantized size clearly distinguished when comparing download sizes or planning local storage.

A parameter-based calculation does not establish how much RAM a particular CPU inference setup will need. Use the estimate as a starting point, then check the intended runtime with the input lengths and summary lengths you expect to use. Avoid turning the weight estimate directly into a claim that the model fits a particular PC.

Apply the same method to another model only after confirming its parameter count and intended weight precision. Where a parameter count is unavailable, leave the memory estimate unspecified. A model name or context limit cannot substitute for the parameter count in this calculation.

Which models provide enough context for your documents?

Google’s google/bigbird-pegasus-large-arxiv offers a 4K-token context window [5], making it the candidate to consider when your documents exceed the 1K-token windows of the Facebook models [1][7] or Google’s google/pegasus-xsum limit of 512 tokens [3].

Facebook’s facebook/bart-large-cnn and Facebook’s facebook/bart-large-xsum both provide 1K tokens of context [1][7]. Context capacity therefore does not distinguish those models. For a document that fits either window, make the choice by comparing summaries against your requirements: retained facts, coverage and readability. Google’s google/pegasus-xsum offers 512 tokens [3]; check whether the intended input fits before adopting it for your workflow.

Measure document length with the chosen model’s tokenizer rather than setting a page-count cutoff. For documents exceeding the selected window, plan to split the text into coherent sections and summarize those separately. Review the combined summary against the original, particularly where an explanation or conclusion depends on material from another section.

For documents beyond 4K tokens, even google/bigbird-pegasus-large-arxiv requires a workflow that handles input outside its stated context window [5]. Treat context capacity as an input constraint, not evidence of summary quality or acceptable CPU performance. Test representative documents on your own machine before committing to a model.

Frequently Asked Questions

How are these CPU-only summarization models ranked?

Ranking note: scope admits only summarization models from labs with a published paper or leaderboard record, using Hugging Face’s summarization pipeline tag. With no public benchmark scoring any candidate, ordering is newest release first, then downloads. The order is facebook/bart-large-cnn from facebook [1], google/pegasus-xsum from Google [3], google/bigbird-pegasus-large-arxiv from Google [5], then facebook/bart-large-xsum from facebook [7]. facebook/bart-large-cnn ranks first because publication dates tie and downloads break the tie, without establishing quality superiority.[1][3][5][7]

Are these actually recent model releases?

No: all candidates were first published on Hugging Face on March 2, 2022.[1][3][5][7] The recent-release pool is too sparse, so the newest available candidates are listed as established alternatives. The established picks by downloads are facebook/bart-large-cnn, google/pegasus-xsum and google/bigbird-pegasus-large-arxiv.[1][3][5] Their recency exemption does not itself award first place; the stated ordering uses publication date and then downloads.

How much memory would quantized weights need?

facebook/bart-large-cnn has 406M parameters [1], giving approximately 203 MB for 4-bit weights, estimated (params x 0.5 bytes).[1] Its published F32 weights occupy 1.6 GB.[1] Treat the calculated value as a weight-only estimate, not a total RAM requirement or confirmed CPU deployment result. A practical memory budget still requires measurement in your chosen runtime.

Which candidate can accept longer documents?

Google’s google/bigbird-pegasus-large-arxiv has the longest listed context here at 4K tokens.[5] Both facebook/bart-large-cnn and facebook/bart-large-xsum list 1K tokens,[1][7] while google/pegasus-xsum lists 512 tokens.[3] Use those limits to screen candidates against your input length. A larger context allowance does not establish better summaries or acceptable CPU latency; verify both with representative documents before choosing a deployment.

Which licenses apply to these models?

facebook/bart-large-cnn and facebook/bart-large-xsum list the MIT license.[1][7] Google’s google/bigbird-pegasus-large-arxiv lists Apache-2.0.[5] The license for google/pegasus-xsum is unspecified here, so leave its licensing status unresolved until you check the repository terms. For a deployment decision, record the applicable license alongside the exact model you intend to distribute or run.

Which model has proven summarization quality or CPU speed?

For facebook/bart-large-cnn,[1] google/pegasus-xsum,[3] google/bigbird-pegasus-large-arxiv,[5] and facebook/bart-large-xsum,[7] the ranking status is “no public benchmark yet.” No candidate can therefore claim a measured quality or CPU-speed advantage in this comparison. Treat the order as a shortlist, then evaluate summary accuracy, omissions, latency and memory on your own documents and hardware before selecting a model.

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