Fastest
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →https://www.patsnap.com/resources/blog/rd-blog/long-context-attention-mechanisms-patent-landscape/ · Patsnap · data cut-off 2026-07-31 · downloaded from the live page
Patent Landscape · AI Inference & ServingLong Context Attention Mechanisms Patents
- Concentrated at the top. The five leading filers together account for 31.5% of all 305 records in scope, with the leader alone holding 49 filings.
- Filings accelerated sharply. Annual filings rose from 19 in 2021 to 67 in 2024, a +253% increase over three years, before the field entered a still-settling 2025-2026 window.
- AI-model classes dominate the claim map. 70.5% of records sit in G06N alone, with G06F close behind at 36.7%, showing where sparse and linear attention claims are being written today.
305
Published Records
31%
Top-5 Share of All Records
+253%
Filing Growth 2021→2024
US
Leading Jurisdiction
Filing growth compares 2021 (19 records) with 2024 (67) — a three-year span. 2024 is the most recent year we treat as complete: publication lags filing by roughly 18 months, so 2025 onwards are still filling in and any growth rate that ends there would understate the field. Top-5 share is the combined record count of the five largest assignees divided by all 305 records in scope (CR5), not by the ranked leaders only.
Published byPatsnap Research··8 min readSourced from Patsnap Eureka
Overview
What this landscape covers
Long context attention mechanisms cover the techniques that let transformer-based models handle sequences far longer than their original training window: sparse attention, sliding window attention, and linear attention, combined with position extrapolation, rotary position embedding, attention masking, context compression and memory-bandwidth optimisation. These are the mechanisms that decide whether a model can serve a million-token context economically or falls back to quadratic-cost attention. The search underlying this page pulls 305 published records filed or published between 2015 and mid-2026 that combine at least one of these attention variants with at least one of these long-context enabling techniques.
Because publication lags filing by roughly 18 months, the most recent filing years understate real activity; the 2025 and 2026 figures will keep rising as more applications publish. The patterns below use patent families as the unit of comparison, which neutralises continuation filings and multi-jurisdiction duplicates.
Filing activity, 2017-2026
Source: Patsnap Eureka. Assignee ranking and totals. Derived from a Patsnap search on Long Context Attention Mechanisms covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
10,000 free credits to start
The Data
Filing trend and technology composition
Filing volume and IPC classification together show where claim activity concentrates and how fast it grew once long-context serving became commercially urgent.
A three-year filing surge, now settling into the publication lag
Annual filings sat near zero through 2017 and climbed slowly until 2021, when 19 records were filed. By 2024 - the most recent year that can be treated as complete - that figure had reached 67, a +253% rise over three years. 2025 shows 113 filings, the peak so far, but because publication trails filing by around 18 months, 2025 and the partial 2026 count of 27 should be read as still filling in rather than as a plateau or decline.
AI-model and general data-processing classes carry the claim weight
G06N (computing arrangements based on specific computational models, i.e. AI/ML) appears in 70.5% of the 305 records in scope, far ahead of any other class, with G06F (general electric digital data processing) present in 36.7%. Image and video classes - G06V, G06T, H04N - each cover a meaningful minority of filings, pointing to attention mechanisms being claimed alongside vision and video pipelines rather than only text. Because records can carry multiple IPC classes, these shares add up to more than 100% of the record total; they are not mutually exclusive buckets.
Shares are the percentage of the 305 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Source: Patsnap Eureka. Filing trend and technology composition. Derived from a Patsnap search on Long Context Attention Mechanisms covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.
Go deeper on Long Context Attention Mechanisms with Eureka
This page is one run against one query. Ask Eureka your own question about long context attention mechanisms and every answer comes back with the patent numbers behind it.
Try EurekaKey Patents
Representative filing and most-cited records
Representative Record
Methods and devices for accelerating a transformer with a sparse attention pattern (US20230133305A1)
BEIJING TRANSTREAMS TECHNOLOGY CO. LTD.
A method and an apparatus for accelerating a transformer with a sparse attention pattern are provided. The method includes that a heterogeneous device including one or more GPUs loads a first matrix, a second matrix, and a transformed sparsity mask into a first sampled dense-dense matrix multiplication (SDDMM) kernel in a sparse attention module in the transformer and generates a first output based on the first matrix, the second matrix, and the transformed sparsity mask by the first SDDMM kernel, generates a second output by a softmax kernel in the sparse attention module based on the first output, loads the second output, a third matrix, and the transformed sparsity mask into a matrix multAbstract truncated as published.
Most-cited records in this landscape
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20220237368A1 | Systems and methods for machine content generation | 269 |
| 2 | US20250259085A1 | Convergent Intelligence Fabric for Multi-Domain Orchestration of Distributed Agents with Hierarchical Memory … | 208 |
| 3 | US20230252224A1 | Systems and methods for machine content generation | 156 |
| 4 | US20210183484A1 | Hierarchical CNN-transformer based machine learning | 72 |
| 5 | US20220318255A1 | Cross-lingual language models and pretraining of cross-lingual language models | 55 |
| 6 | US20230351102A1 | Machine content generation | 52 |
| 7 | US20250390352A1 | AI Serving Hardware and Software Frontier Enhancements | 39 |
| 8 | US11748555B2 | Systems and methods for machine content generation | 36 |
| 9 | US20220383048A1 | Systems and methods for cross-lingual cross-modal training for multimodal retrieval | 33 |
| 10 | US20240112088A1 | Vector-Quantized Image Modeling | 32 |
Citation counts favour older records simply because they have had more time to accumulate citations inside the searched corpus; treat them as a signal of influence, not of current technical importance.
Each row carries its publication number; clicking a row searches Eureka by that number.
Source: Patsnap Eureka. Citation counts and representative records. Derived from a Patsnap search on Long Context Attention Mechanisms covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Run it yourself
Put your own technology through the same analysis
Where to run it
For builders
MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →Insights
What the numbers mean for filing strategy
Three patterns matter more than the raw counts: where filing concentrates, how fast the field moved, and which technical routes are actually being claimed together.
Concentration
31.5%
of 305 records held by the top 5 filers
A short leader group, then a long tail
The five leading assignees together hold 31.5% of all 305 records in scope, and the leader alone accounts for 49 filings - well ahead of the fifth-place holder at 11. Below the top ten, filing drops to single- and double-digit counts across the remaining ranked entrants, a classic long-tail pattern where most organisations have filed once or twice.
Based on the 100-company assignee ranking returned by the dataset.
Growth
+253%
filing growth, 2021 to 2024
Activity accelerated sharply through 2024
Annual filings grew from 19 in 2021 to 67 in 2024, a +253% increase over three years, coinciding with the commercial push toward longer context windows in deployed language models. 2025's count of 113 is the highest on record but, given the roughly 18-month publication lag, should not yet be read as the field's ceiling.
2024 is the most recent year treated as complete for growth comparisons.
Composition
70.5%
of records classified under G06N
Claims cluster on the AI-model layer
G06N covers 70.5% of the 305 records, meaning most filings frame their contribution as a computational model or learning architecture rather than as generic data processing (G06F, 36.7%) or as a media-specific application. Filers targeting vision or audio pipelines (G06V, G06T, G10L) sit in smaller but non-trivial minorities.
Class shares sum to more than 100% because records carry multiple IPC codes.
Jurisdictions
138
US-received filings, the largest single office
Filing is US-centred with a strong PCT track
The United States receives more filings in this dataset than any other office (138), with WIPO PCT filings (59) indicating a substantial share of applicants pursuing multi-jurisdiction protection from the outset. China (38), India (27), Europe (19) and Australia (7) round out the remaining filing activity.
Counts are by receiving office, not by applicant nationality.
Eureka AI Agent
Looking for what nobody has claimed yet?
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to long context attention mechanisms, with the prior art for and against each one.
Source: Patsnap Eureka. Co-assignee relationships and derived observations. Derived from a Patsnap search on Long Context Attention Mechanisms covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Players
Who is filing, and where momentum is shifting
The leader board is short and dominated by large platform and chip vendors, but recent-year momentum data suggests most of the visible leaders slowed their filing pace in the latest tracked year - consistent with the publication lag rather than a genuine pullback.
Leader
49
records held by the top-ranked assignee
One filer well ahead of the pack
The leading assignee's 49 records put it far ahead of the rest of the field; fifth place holds 11 and tenth place holds 8, so the drop-off from first to fifth is steep before the ranking flattens into a longer tail.
Ranking is based on patent family counts, not raw document counts.
Momentum
-67% to -100%
YoY change among several leading filers
Latest-year counts look soft across the board
Several of the most active historical filers show sharp year-over-year declines in the latest tracked year, including drops to zero for some. Given the roughly 18-month gap between filing and publication, this reads as an artefact of the data cut-off rather than as leaders exiting the field.
Momentum figures compare the latest tracked year against the prior year.
Collaboration
10
co-assignee pairs identified
Co-filing is limited but concentrated
Only 10 co-assignee pairs appear in the dataset, with the strongest pairings linking a major GPU vendor to a research university and to a medical-technology partner, and a major electronics manufacturer to its own China-based research subsidiary. Collaboration is the exception here, not the norm.
Pair counts reflect records naming both assignees jointly.
Recent-year filing momentum by assignee
| Assignee | Recent year | YoY |
|---|---|---|
| Google LLC | 2 | -67% |
| GDM Holding LLC | 1 | -80% |
| Qualcomm Inc. | 0 | -100% |
| NVIDIA Corp. | 0 | -100% |
| ATTENTION LABS INC | 0 | -100% |
| Microsoft Technology Licensing, LLC | 0 | -100% |
| Samsung Electronics Co., Ltd. | 0 | -100% |
| Oracle International Corp. | 0 | -100% |
Source: Patsnap Eureka. Assignee-level momentum. Derived from a Patsnap search on Long Context Attention Mechanisms covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
What's Next
Where to take this analysis
The dataset points to specific next steps depending on whether the goal is freedom-to-operate, portfolio positioning, or scouting for partners.
Check freedom-to-operate against the top filers
With 31.5% of records held by five assignees, any product roadmap touching sparse attention, sliding-window attention or linear attention should be checked against those portfolios before committing engineering time.
Run a freedom-to-operate check in EurekaTrack the settling 2025-2026 filing window
The apparent 2025 peak and 2026 partial count will keep shifting as filings publish; re-running this search periodically will show whether growth continued past the +253% (2021-2024) rate or slowed.
Set up filing alerts in EurekaScout the under-claimed branches
KV-cache compression, runtime-adaptive sparsity and hardware-aware kernel scheduling show thinner filing density than the core mechanisms and may offer clearer claim space for new filings.
Explore white space in EurekaSource: Patsnap Eureka. Forward-looking reading of the same dataset. Derived from a Patsnap search on Long Context Attention Mechanisms covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
FAQ
Common questions about this landscape
The dataset's assignee ranking shows a single leader holding 49 of the 305 records in scope, well ahead of the rest of the ranked entrants. The top five assignees combined hold 31.5% of all records, and the top ten hold 45.2%, indicating that a small group of large technology and chip companies files far more than the median entrant. Below tenth place, filing counts drop quickly into single digits, forming a long tail of organisations with only one or two filings each.
Yes, through the last complete year of data. Annual filings rose from 19 in 2021 to 67 in 2024, a +253% increase over three years, reflecting the industry's push toward longer context windows in deployed models. Figures for 2025 and 2026 appear higher still, but because publication typically lags filing by about 18 months, those years are undercounted and should not be read as a plateau or a slowdown.
G06N covers computing arrangements based on specific computational models, which in practice means machine-learning and AI architectures. That class appears in 70.5% of the 305 records in scope, meaning most applicants are framing their attention-mechanism inventions as model-level or algorithmic contributions rather than as generic data-processing improvements. G06F, the general digital-data-processing class, appears in 36.7% of records, often alongside G06N on the same filing since a record can carry multiple classes.
Filing density is comparatively thin in branches adjacent to the core sparse, sliding-window and linear attention mechanisms, including KV-cache compression for streaming inference, hardware-aware attention kernel scheduling, cross-modal long-context fusion, and runtime-adaptive sparsity pattern selection. These sit next to well-covered core mechanisms but have not attracted the same concentration of filings from the leading assignees, which typically means clearer room to file a first claim without colliding with dense prior art.
The United States receives the largest share of filings in this dataset at 138 records, followed by WIPO's PCT route at 59, which signals a substantial portion of applicants seeking multi-jurisdiction coverage from a single international filing. China follows with 38, then India with 27, Europe with 19, and Australia with 7. These are counts by receiving office rather than by the applicant's home country, so a US-headquartered filer's PCT application would count toward the WIPO figure, not necessarily the US one.
Answers are grounded in the same dataset. Derived from a Patsnap search on Long Context Attention Mechanisms covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Research Long Context Attention Mechanisms in depth with Eureka
Go past this page: query the whole long context attention mechanisms corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company's registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.