As of November 17, 2026, cross-motions for summary judgment on infringement and fair use, filed June 29, 2026, are pending before Judge Charles R. Breyer, with no ruling located, no certified class and no trial date in the record.
The case began as O'Nan v. Databricks, filed on March 8, 2024, and was consolidated with Makkai v. Databricks. The court dismissed the claim over the DBRX model in August 2025 and let a reworked version proceed in April 2026.
DataPacts is not affiliated with any party, and this page is information, not legal advice, within the independent reference on AI training data licensing.
Dates, docket numbers and quotations come from the court filings named in each table, read on November 17, 2026, and how figures are sourced and dated is set out on the methodology page.
In re Mosaic LLM Litigation at a Glance: Court, Docket, Parties and Status
The case is in the summary judgment stage, with no merits ruling, no class certification order and no trial date in the record. The table lists the case facts as of 2026-11-17.
| Field | Value | Source |
|---|---|---|
| Case | In re Mosaic LLM Litigation (formerly O'Nan et al. v. Databricks, Inc. et al.) | Dkt. 77, 78 |
| Court | U.S. District Court, N.D. Cal., San Francisco Division; Judge Charles R. Breyer; discovery referred to Magistrate Judge Lisa J. Cisneros | Dkt. 249, 471 |
| Master docket | 3:24-cv-01451-CRB, consolidated with 3:24-cv-02653-CRB (Makkai) | Dkt. 471 |
| First complaint filed | 2024-03-08, O'Nan, Nazemian and Keene v. Databricks and Mosaic ML | Dkt. 1 |
| Plaintiffs | Stewart O'Nan, Abdi Nazemian, Brian Keene, Rebecca Makkai, Jason Reynolds (proposed class) | Dkt. 324 |
| Defendants | Databricks, Inc.; Mosaic ML, LLC, formerly Mosaic ML, Inc. | Dkt. 39, 321 |
| Claims | 4 counts: direct, vicarious, contributory infringement and inducement (17 U.S.C. 501) | Dkt. 254 |
| Models named | MPT-7B, MPT-30B, MPT-7B-StoryWriter (MosaicML); DBRX (Databricks) | Dkt. 1, 162, 254 |
| Collections named | RedPajama-Books, Books3, The Pile (alleged) | Dkt. 1, 287 |
| Status | cross-motions for summary judgment filed 2026-06-29, noticed for hearing 2026-10-30; no ruling located; no class certified; no trial date in the record | Dkt. 290, 321, 324, 471 |
What the Authors Claim: Their Books in MosaicML's MPT Training Data
The authors claim that MosaicML trained its MPT models on a books collection that contained their registered works, and that Databricks is liable as MosaicML's corporate parent and, since 2026, directly for its DBRX work.
The original complaint named MPT-7B and MPT-30B (Dkt. 1), and the complaint and the plaintiffs' motion date the Databricks acquisition of MosaicML to July 2023.
The court itself calls Databricks "the corporate parent of Mosaic ML" (Dkt. 129). Every statement about books and datasets below is the authors' allegation, and no court has made a finding on any of it.
How the Authors Trace Their Books Into the MPT Training Mix
The complaint builds a chain from MosaicML's own MPT-7B announcement to the authors' books: the announcement lists RedPajama-Books in the training mix, that slice is alleged to be a copy of Books3, and Books3 is alleged to contain the works in Exhibit A. The announcement describes a "MosaicML-curated mix of sources" that "includes elements of the recently-released RedPajama dataset" (Dkt. 1, paras. 24 to 26).
The page on RedPajama and its book slice describes what the corpus contains and why the slice is marked defunct.
The Second Amended Consolidated Complaint adds that MosaicML obtained RedPajama in April 2023. It quotes MosaicML on MPT-7B-StoryWriter as "further trained on a filtered fiction subset of the [B]ooks3 dataset" (Dkt. 254, paras. 39 and 46). Much of that complaint is redacted, so this page states only the public text.
| Collection | What the authors allege | Source |
|---|---|---|
| RedPajama-Books | listed in MosaicML's MPT-7B training-mix table; obtained April 2023; its books component is a copy of Books3 | Dkt. 1 paras. 24 to 26; Dkt. 254 paras. 39, 45 |
| Books3 | derived from the Bibliotik private tracker; contains the Exhibit A works; removed from Hugging Face in October 2023; a filtered fiction subset used for MPT-7B-StoryWriter | Dkt. 1 paras. 27 to 31; Dkt. 254 para. 46 |
| The Pile | used with RedPajama-Books in MPT2 experiments, a model later renamed DBRX | Dkt. 287, summarizing Dkt. 254 para. 52 |
The table compares the three collections as alleged in the operative complaint of 2026-01-21; no court has made a finding on any row.
Books3 is alleged to derive from the Bibliotik private tracker and to have been taken down from Hugging Face in October 2023 as "defunct". The page on Books3, the collection the authors name covers its origin, counts and removal, and this page adds nothing about the collection beyond the allegations.
Four Counts, Two Defendants and the Proposed Class
The operative complaint, the Second Amended Consolidated Complaint filed January 21, 2026 (Dkt. 254), pleads four counts:
- Count 1: direct copyright infringement under 17 U.S.C. 501, against all defendants;
- Count 2: vicarious infringement, against Databricks;
- Count 3: contributory infringement, against all defendants;
- Count 4: inducement of infringement, against all defendants.
The proposed class covers legal or beneficial owners of works that the defendants downloaded, reproduced or distributed, where the work carries an ISBN or ASIN and was registered within five years of first publication or within three months of publication. No class is certified in the record read on November 17, 2026.
The DBRX Claim Against Databricks: Dismissed in 2025, Reinstated in 2026
The court dismissed the authors' first direct infringement claim against Databricks over its DBRX models in August 2025 and let a reworked version proceed in April 2026.
On May 20, 2025 the magistrate judge allowed discovery on how DBRX was trained, noting that the defendants said it was trained "using a different mix of data" (Dkt. 123).
On June 25, 2025 the court granted leave to add a direct claim over DBRX (Dkt. 129), and on August 19, 2025 it dismissed that claim without leave to amend (Dkt. 162).
The statement that DBRX was trained on "open data sets that the community knows" was too generalized, unlike the RedPajama-Books allegations against MosaicML, which "state a claim".
After discovery the court allowed a second amendment on January 20, 2026, calling good cause "a close call" (Dkt. 251). On April 21, 2026 it denied the defendants' motions to dismiss and strike (Dkt. 287).
The new allegations describe MPT2 experiments "using The Pile and RedPajama-Books", a model later renamed DBRX; the plaintiffs argued that copying is copying at any stage of training, and the court found the allegations sufficient for now. The page on The Pile covers EleutherAI's mixed corpus.
The court limited its own holding: "Defendants may ultimately prevail on this issue, but for now, Plaintiffs' allegations are sufficient." The defendants answered the Second Amended Consolidated Complaint on May 11, 2026 (Dkt. 292).
In re Mosaic Docket Entries, 2024 to 2026
The docket runs from the March 2024 complaint through consolidation, the DBRX dismissal and reinstatement and the 2026 summary judgment briefing to a plaintiffs' notice of October 2, 2026, the latest entry located. The table lists 17 entries as of 2026-11-17.
| Date | Dkt. | Entry | What it decided or recorded | Source |
|---|---|---|---|---|
| 2024-03-08 | 1 | O'Nan complaint, class action, jury demand | MPT trained on RedPajama-Books, alleged copy of Books3 | RECAP archive |
| 2024-05-13 | 45 | relation order | Makkai v. Databricks (No. 3:24-cv-02653) related to O'Nan | RECAP archive |
| 2024-12-12 | 57 | consolidation order | cases consolidated on the parties' stipulation (as recited in Dkt. 73) | RECAP archive |
| 2025-05-20 | 123 | discovery order (magistrate judge) | discovery on how DBRX was trained allowed | RECAP archive |
| 2025-06-25 | 129 | leave to amend | DBRX direct claim and updated works list allowed | RECAP archive |
| 2025-08-08 | 152 | subpoena order (magistrate judge) | request to nonparty Microsoft granted in part | RECAP archive |
| 2025-08-19 | 162 | order granting motion to dismiss | DBRX direct claim dismissed without leave to amend | RECAP archive |
| 2026-01-05 | 242 | discovery order (magistrate judge) | acquisition agreement and retention policies not ordered | RECAP archive |
| 2026-01-16 | 249 | discovery order (magistrate judge) | executed text-data agreements due 2026-01-30 | RECAP archive |
| 2026-01-20 | 251 | leave to amend | direct, contributory and inducement claims; new class definition | RECAP archive |
| 2026-01-21 | 254 | Second Amended Consolidated Complaint | four counts | RECAP archive |
| 2026-04-21 | 287 | order denying motions to dismiss and strike | DBRX claim proceeds | RECAP archive |
| 2026-05-11 | 292 | defendants' answer | answer to the Second Amended Consolidated Complaint | RECAP archive |
| 2026-06-29 | 321 | defendants' motion for summary judgment | all claims; fair use | RECAP archive |
| 2026-06-29 | 324 | plaintiffs' partial motion for summary judgment | downloading and distribution as infringement | RECAP archive |
| 2026-09-14 | 468 | plaintiffs' reply | briefing on the cross-motions complete under the schedule | RECAP archive |
| 2026-10-02 | 471 | plaintiffs' statement of recent decision | cites a Third Circuit opinion, No. 25-2153 | RECAP archive |
Entries come from the RECAP archive of the court docket; no later entry was located on 2026-11-17.
The Pending Cross-Motions for Summary Judgment
Both sides moved for summary judgment on June 29, 2026, and the court has not ruled as of November 17, 2026.
The defendants' motion (Dkt. 321) seeks judgment on all claims and presents four questions: whether using the books as training data for LLM research and development is fair use; whether giving nine customers and a graduate student access to datasets that contain the books is fair use; whether there is no secondary liability; and whether DBRX was trained on the books at all.
In it the defendants state that they "used Plaintiffs' books, along with hundreds of billions of words of other text, to train a series of open-source LLMs and conduct LLM research", and they argue fair use. That is the defendants' own statement in a brief, not a finding.
The plaintiffs' partial motion (Dkt. 324, cited later as ECF No. 378) targets two acts as infringement apart from training: downloading the books from an unauthorized source to avoid paying a purchase price, and reproducing and distributing them through a commercial library on public cloud storage.
A plaintiffs' filing of July 20, 2026 (Dkt. 379) states that the defendants do not contest ownership of the asserted works.
The schedule set oppositions for August 13, 2026 and replies for September 14, 2026, with the hearing noticed for October 30, 2026 (Dkt. 290).
Amici filed motions for leave on August 27, 2026: copyright law professors and the Electronic Frontier Foundation with Chamber of Progress support the defendants (Dkt. 422, 443), and the RIAA with six music industry groups oppose them (Dkt. 430).
The defendants also moved to exclude testimony of a plaintiffs' expert, with a reply on September 14, 2026 (Dkt. 463). No order deciding any of these motions was located.
The doctrine itself and other courts' rulings sit on the page on fair use and AI training in US law.
Discovery Orders on Training-Data License Agreements
The magistrate judge ordered Databricks and MosaicML to produce their executed agreements for textual training data by January 30, 2026, while holding that related communications need not be produced (Dkt. 249, January 16, 2026). The request covered agreements with third parties for curated LLM training data.
The order says that what the defendants paid for curated datasets used to fine-tune models "says relatively little" about whether, or how much, they would have paid the authors, yet it orders production because fair use in AI training is novel.
An earlier order of January 5, 2026 denied the plaintiffs' requests for the MosaicML and Databricks acquisition agreement and for retention policies, while granting a third set of production requests in part (Dkt. 242).
On August 8, 2025 the magistrate judge granted in part the plaintiffs' subpoena to the nonparty Microsoft (Dkt. 152).
As Dkt. 249 describes it, Microsoft's licenses with publishers to obtain books for LLM training were found relevant to fair use, because they could show what licensing would have paid. The agreements are not on the public docket.
What the Mosaic Record Shows Holders of Book Rights
The record shows that courts in this case treat the route by which books entered a training set as a separate question from training itself. The plaintiffs' partial motion concerns downloading and distribution, not training, while the defendants frame research and training as fair use.
The wider practice of copying from unauthorized collections is covered on the page on pirated books and shadow libraries in AI training, which gathers what courts have found about pirated copies.
The license-agreement orders treat paid book licenses as possible evidence of a market for the works, with no price from this case.
The licensed route has its own page on book publishers and authors licensing books for AI training. Not licensing, and waiting for rulings, are also choices open to a holder.
Where In re Mosaic Sits Among Authors' Book Cases
In re Mosaic is one of the pending authors' class actions over book collections in training data with summary judgment motions fully briefed.
The case tracker of AI copyright lawsuits over training data lists every tracked case with its docket and status. The table compares five related cases as of 2026-11-17, each with the status its own page records.
| Case | Court and docket | Books or collection at issue | Status | Source |
|---|---|---|---|---|
| Bartz v. Anthropic | N.D. Cal. 3:24-cv-05417-AMO | Books3, LibGen and PiLiMi copies | settled; final approval 2026-07-20 (Dkt. 680) | Court docket |
| Huckabee v. Bloomberg | S.D.N.Y. 1:23-cv-09152-MMG | Books3 for BloombergGPT | discovery; deadlines extended 90 days 2026-11-05 (Dkt. 139), as of 2026-11-17 | Court docket |
| In re Salesforce | N.D. Cal. 3:25-cv-08862-CRB | Books3 via The Pile and RedPajama for XGen | pending, in discovery | Court docket |
| Kadrey v. Meta | N.D. Cal. 3:23-cv-03417-VC | shadow-library books for Llama | training claim decided for Meta 2025-06-25; first coordinated trial set for 2027-05-03 | Court docket |
| Nazemian v. NVIDIA | N.D. Cal. 4:24-cv-01454-JST | Books3 for NeMo Megatron | discovery, as of 2026-11-17 | Court docket |
Three of the original authors here, O'Nan, Nazemian and Keene, filed Nazemian v. NVIDIA on the same day, March 8, 2024, and In re Salesforce is before the same court.
More About In re Mosaic LLM Litigation
This page is information, not legal advice.
Decided Versus Still Open in In re Mosaic
The court has decided the pleadings and several discovery disputes; it has not decided infringement, fair use, class certification or damages.
Decided:
- the two complaints were related and consolidated (Dkt. 45, 57);
- the DBRX claim was dismissed in 2025 and reinstated in 2026 (Dkt. 162, 287);
- the defendants' text-data agreements were ordered produced (Dkt. 249);
- the request for the acquisition agreement was denied (Dkt. 242).
Still open:
- both summary judgment motions (Dkt. 321, 324);
- the motion to exclude a plaintiffs' expert (Dkt. 463);
- class certification;
- damages, settlement and trial, none of which appears in the record.
Frequently Asked Questions About In re Mosaic LLM Litigation
Is In re Mosaic LLM Litigation the same case as O'Nan v. Databricks?
Yes: O'Nan (No. 3:24-cv-01451) and Makkai (No. 3:24-cv-02653) were related in May 2024, consolidated in December 2024 and recaptioned In re Mosaic LLM Litigation in 2025. The recaption was filed on January 16, 2025 (Dkt. 77) and used from February 12, 2025 (Dkt. 78).
Has the court ruled on fair use in the Mosaic case?
No: both summary judgment motions were filed on June 29, 2026 and noticed for October 30, 2026, and no ruling was located as of November 17, 2026.
Is there a settlement or a trial date?
Neither appears in the record read as of November 17, 2026. The latest entry located is Dkt. 471 of October 2, 2026.
Who is in the proposed class?
The proposed class covers owners of registered works that the defendants allegedly downloaded, reproduced or distributed, with an ISBN or ASIN. No class has been certified (Dkt. 254).
Is the case about Databricks Marketplace?
No: the case concerns the MPT and DBRX models, not a data marketplace. The Databricks profile covers how the marketplace lists data.