Model Quantization Patents: Who Leads, Where the Gaps Are 2026
Filing growth compares 2021 (1,548 records) with 2024 (1,833) — 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 263,013 records in scope (CR5), not by the ranked leaders only.
What the model quantization patent record actually shows
Model quantization — compressing neural network weights and activations to lower-bit representations without unacceptable accuracy loss — sits at the intersection of AI model design and classic signal-coding engineering. The 263,013 records in scope span 2015 through mid-2026, covering everything from generic quantized model claims to video-codec-specific bit-depth reduction. The dataset’s IPC spread is telling: pictorial communication (H04N) and coding conversion (H03M) subclasses carry more filings than the dedicated AI-model computing class (G06N), which means a meaningful share of ‘quantization’ patents originate from video and telecom engineering groups rather than pure machine-learning labs.
Filing activity rose steadily through the period tracked, peaking so far in 2024 at 1,833 records, with 2025 and 2026 figures still understated because publication trails filing by roughly 18 months. Assignee concentration is moderate rather than extreme: the leading company holds 22,630 records, but the top 10 combined account for just 30.8% of all records in scope, leaving substantial room for new entrants and narrower technical claims.
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
Filing trends and technology composition
These figures come directly from the 263,013 records currently in scope for model quantization filings, spanning original applications and their published family members from 2015 to mid-2026.
A decade of steady growth, understated at the edges
Annual filings moved from 484 in 2017 to a peak of 1,833 in 2024, a +18% rise over the 2021-2024 window alone. The apparent drop-off into 2025 and 2026 reflects publication lag rather than a real slowdown — most 2025-26 filings have not yet published.
Where quantization claims actually sit
H04N (pictorial communication) leads at 3.4% of all records, ahead of G06N (AI-model computing) at 1.6%. Because records often carry multiple IPC codes, these shares are not mutually exclusive, but the ranking itself is informative: quantization claims are as often filed as video-coding or speech-processing improvements as they are filed as general AI-model patents.
Shares are the percentage of the 263,013 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Foundation Models: Model Quantization Patent Landscape with Eureka
This page is one run against one query. Ask Eureka your own question about foundation models: model quantization patent landscape and every answer comes back with the patent numbers behind it.
Try EurekaA representative claim and the most-cited prior art
US20230394289A1 — Learning model quantization program and method
A non-transitory computer-readable recording medium stores a learning model quantization program for causing a computer to execute a process including: in an objective function for searching for a combination of layers in which parameters of a machine-learned model using a neural network are quantized, the objective function including inference accuracy of the quantized model and an index related to a compression ratio of the model, setting a specific gravity such that the specific gravity of the index related to the compression ratio with respect to the inference accuracy decreases as the compression ratio increases; selecting a layer in which the objective function is optimized, as a layer.Filed by Fujitsu, published 2023-12-07.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US6839682B1 | Predictive modeling of consumer financial behavior using supervised segmentation and nearest-neighbor matching | 866 |
| 2 | US5889868A | Optimization methods for the insertion, protection, and detection of digital watermarks in digitized data | 846 |
| 3 | US20170332095A1 | Affine motion prediction for video coding | 794 |
| 4 | US5027406A | Method for interactive speech recognition and training | 683 |
| 5 | US7003035B2 | Video coding methods and apparatuses | 621 |
| 6 | US6072496A | Method and system for capturing and representing 3D geometry, color and shading of facial expressions and oth… | 580 |
| 7 | US20050123207A1 | Video frame or picture encoding and decoding | 570 |
| 8 | US5231484A | Motion video compression system with adaptive bit allocation and quantization | 552 |
| 9 | US6061793A | Method and apparatus for embedding data, including watermarks, in human perceptible sounds | 536 |
| 10 | US20130155058A1 | Four-dimensional augmented reality models for interactive visualization and automated construction progress m… | 520 |
Citation counts favor older filings simply by virtue of being searchable longer; treat them as a signal of influence within this corpus, not as a ranking of current technical importance.
Each row carries its publication number; clicking a row searches Eureka by that number.
Put your own technology through the same analysis
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 →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 →What the data means for a filing or freedom-to-operate decision
Three patterns in this corpus matter more than the raw counts: where filings cluster by technology class, how concentrated ownership is at the top, and how recent activity by the biggest names is trending.
Leadership is real but not a wall
The top 5 of 100 ranked assignees combine for 54,513 records, 20.7% of everything in scope. That leaves roughly four-fifths of the corpus held outside the leading cluster — dense in specific claim areas, but not a field one company can block outright.
Video coding outweighs generic AI-model claims
H04N pictorial communication carries more quantization filings than G06N AI-model computing. Anyone assuming this is purely a machine-learning patent space will miss the video- and speech-codec prior art that actually crowds core compression claims.
Growth held through the last complete filing year
Filings rose from 1,548 in 2021 to 1,833 in 2024. That is the most recent year that can be read reliably; 2025-26 figures will fill in as publications catch up, so a reader should not treat the current tail as a slowdown.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to foundation models: model quantization patent landscape, with the prior art for and against each one.
Turning this landscape into a filing or freedom-to-operate position
The counts and rankings on this page describe where claims already sit. Deciding where to file next, or whether a specific compression technique is clear to build, needs a closer read of the actual claim language behind the numbers.
Map claim scope, not just counts
A record count tells you density, not what is actually claimed. Reading the independent claims behind the leading assignees' filings in your specific sub-area — mixed-precision allocation, calibration methods, hardware-adaptive selection — shows what is actually blocked versus merely adjacent.
Explore claims in EurekaTrack the momentum shift as 2025-26 data fills in
Current year-over-year figures look like a drop for every major assignee, but that is a publication-lag artifact. Revisiting this view in a future cut, once 2025 filings finish publishing, will show whether the +18% growth trend into 2024 continued.
Set up ongoing monitoring in EurekaTest a draft claim against the white space
The under-claimed branches identified here — post-training calibration, outlier-aware clipping, and similar — are starting points, not guarantees. Running a candidate claim against the full corpus before drafting avoids discovering the overlap after filing.
Check novelty in EurekaCommon questions about model quantization patents
The dataset ranks 100 assignees by record count, with the leading company holding 22,630 of the 263,013 records in scope. The top five combined account for 54,513 records, 20.7% of the total, which is meaningful concentration but far from exclusive control. Fifth place holds 6,488 records and tenth place holds 4,143, showing a steep drop after the very top few filers rather than a smooth distribution. Anyone doing freedom-to-operate work should check the leading handful closely but not assume the field stops there.
Yes, through the last complete filing year the trend is upward: filings rose from 1,548 in 2021 to a peak of 1,833 in 2024, an 18% increase over that span. Figures for 2025 and 2026 appear lower, but that reflects publication lag of roughly 18 months rather than a real slowdown, since many recent filings have not yet published. Treat the most recent one to two years in any chart as provisional and likely to be revised upward as more records surface.
The IPC composition shows quantization claims spread well beyond generic AI-model computing. H04N (pictorial communication, largely video and TV coding) is the single largest class at 3.4% of all records, ahead of G06N (AI-model computing) at 1.6%, with G10L (speech and audio), G06T (image processing) and H03M (coding and code conversion) also carrying substantial shares. This means a search limited to AI-specific classes will miss a large share of relevant prior art sitting in video-codec and signal-coding filings.
This Fujitsu filing, published 2023-12-07, covers a computer-readable medium storing a program that quantizes a machine-learned neural network by optimizing an objective function balancing inference accuracy against compression ratio, where the weighting given to compression decreases as the compression ratio itself increases. It selects which layers to quantize based on where that objective function is optimized. The claim is specific to this adaptive-weighting, layer-selection mechanism rather than to quantization in general, so it constrains a particular calibration strategy more than the broader technique.
Based on the IPC composition, sub-areas like post-training calibration methods, mixed-precision layer allocation, hardware-adaptive bit-width selection, and outlier-aware weight clipping sit outside the densest classes (H04N, G06N, G10L) and appear less crowded relative to core compression claims. These are not guaranteed-open areas, but they carry lighter filing density in this corpus and are worth a closer novelty check before drafting. The steep drop in record counts after the top few assignees also suggests room for narrowly scoped claims even within crowded IPC classes.
Research Foundation Models: Model Quantization Patent Landscape in depth with Eureka
Go past this page: query the whole foundation models: model quantization patent landscape 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.