Gemma 4 31B is now available in AIP through AWS Bedrock for eligible commercial enrollments.

Model overviews

Gemma 4 31B is Google's 30-billion parameter dense open weights model supporting multimodal input across text and image, featuring a 256K token context window. Gemma 4 31B is available for commercial enrollments with Google | Bedrock enabled in Control Panel.

Getting started

To use these models:

• Confirm that your enrollment administrator has enabled the relevant model family.
• Review token costs and pricing.
• See the complete list of models available in AIP.

Your feedback matters

We want to hear about your experiences using language models in the Palantir platform and welcome your feedback. Share your thoughts with Palantir Support channels or on our Developer Community using the language-model-service tag.

Gemini 3.7 Flash is now available in AIP for non-georestricted, US georestricted, EU georestricted, IL2, IL4, and IL5 enrollments with Google Vertex AI enabled.

Model overview

Gemini 3.7 Flash is the next iteration in the Gemini 3 model family, featuring algorithmic improvements to its core reasoning foundation. For more information, review Google's Gemini 3.7 Flash model card and Google's model announcement.

Availability

Gemini 3.7 Flash is available on:

• Non-georestricted commercial enrollments with Google Vertex AI enabled
• US georestricted commercial enrollments with Google Vertex AI enabled
• EU georestricted commercial enrollments with Google Vertex AI enabled
• IL2, IL4, and IL5 enrollments with Gemini through Google Vertex AI enabled

Getting started

To use this model:

• Confirm that your enrollment administrator has enabled the xAI model family.
• Review token costs and pricing.
• See the complete list of models available in AIP.

Your feedback matters

We want to hear about your experiences using language models in the Palantir platform and welcome your feedback. Share your thoughts with Palantir Support channels or on our Developer Community using the language-model-service tag.

Grok 4.6 is now available in AIP on US georestricted and non-georestricted enrollments with xAI enabled.

Model overview

Grok 4.6 is xAI's latest model, optimized for long-running agents, complex coding and knowledge work, and ambitious interactive or visual projects. For more information, review xAI's Grok 4.6 announcement.

Model metadata

• Context window: 500,000 tokens
• Modalities: Text and image input; text output
• Capabilities: Structured outputs, reasoning, and tool calling

Getting started

To use this model:

• Confirm that your enrollment administrator has enabled the xAI model family.
• Review token costs and pricing.
• See the complete list of models available in AIP.

Your feedback matters

We want to hear about your experiences using language models in the Palantir platform and welcome your feedback. Share your thoughts with Palantir Support channels or on our Developer Community using the language-model-service tag.

Working with media in functions is now generally available. Functions can accept media items as inputs, upload new media, attach media to Ontology objects through Ontology edits and function-backed actions, and read both the bytes and the metadata of any media item. These capabilities are available for TypeScript v2, Python, and TypeScript v1 functions.

A function reads media from media sets; this means that functions can use documents, images, audio, and video that have already been ingested without having to export them somewhere else.

Capabilities of media in functions

• Accept and return media: Take a Media as an input or return a Media as an output, or read Media from a media reference property on an Ontology object.
• Upload media from inside a function: Turn raw bytes into a Media with uploadMedia in TypeScript v2 or client.ontology.media.upload_media in Python. Python also has an awaitable async_upload_media.
• Attach media to Ontology objects: Set a media property when you create or modify an object in an Ontology edit function, or pass a media reference in as an action parameter.
• Read bytes and metadata: Call fetchContents and fetchMetadata in TypeScript v2, or get_media_content and get_media_metadata in Python. Python's get_media_full_metadata and TypeScript v2's fetchFullMetadata return format-specific fields such as page counts for documents or dimensions and bands for imagery.
• Transform media: Rotate and re-encode images, run OCR on a document, render a PDF page as an image, slice a page range into a new PDF, or draw annotations onto an image.
• Select media in live preview: When you live-preview a function that takes a media parameter, you can now search your media sets and choose an item inline instead of manually pasting in a media reference.

Working with media in a function

TypeScript v2 and Python both use the Media type, which wraps a media reference and adds higher-level operations for reading contents and metadata and for attaching the media to an object. The examples below uploads raw bytes and returns the resulting Media:

import type { Client, Media } from "@osdk/client";
import { uploadMedia } from "@osdk/functions";

// Typescript V2
export default async function uploadMediaItem(
    client: Client,
    body: string,
    fileName: string,
): Promise<Media> {
    const blob = new Blob([body], { type: "text/plain" });
    const media: Media = await uploadMedia(client, { data: blob, fileName });
    return media;
}

Python

from ontology_sdk import FoundryClient
from foundry_sdk_runtime.media import Media
from functions.api import function

@function
def upload_media(body: str, media_set_filename: str) -> Media:
    client = FoundryClient()
    media: Media = client.ontology.media.upload_media(
        body=body.encode("utf8"),
        filename=media_set_filename,
    )
    return media

Uploaded media is temporary until you set it on an Ontology object’s media reference property. When the Ontology edits are applied, the media persists on that property. Review the media documentation for the full set of examples, including Ontology edits and Action parameters. TypeScript v1 functions continue to use the MediaItem type, which has built-in operations for document text extraction, OCR, and audio transcription. Review the TypeScript v1 media documentation.

Media transformations

A transformation derives new media from existing media. You can rotate or re-encode an image, render a PDF page as a PNG, slice a page range into a new PDF, run OCR to recover text along with per-word bounding boxes, or draw annotations onto an image. The SDK submits the transformation job, polls it to completion, and returns the bytes to your function, so a single function can chain several steps: render each page, OCR that same page, then annotate the render with the boxes the OCR found.

Because transformations are in beta, call the SDK helpers rather than the media transformation endpoints directly. Use transform_and_wait or async_transform_and_wait on the generated Python client, and transformAndWait in TypeScript v2 (@osdk/api 2.55.0 or greater). The helpers keep transformation job details out of your function, so if those details change while the feature is in beta, your code does not have to change with them.The documentation has a worked example for each of these:

• Run page-by-page OCR on a PDF with bounding box output
• Render PDF pages as images and slice page ranges
• Annotate every page with detected bounding boxes

OCR on dense pages can exceed the default function execution timeout. See Manage published functions to configure a longer one.

Get started

Review the media in functions documentation for TypeScript v2, Python, and TypeScript v1 examples of the workflows described above.

Your feedback matters

We want to hear about your experience building functions over media and welcome your feedback. Share your thoughts with Palantir Support channels or on our Developer Community using the functions or media-sets tag.

You can now configure a custom compute profile in addition to the five standard profile sizes in faster pipelines backed by DataFusion. Custom profiles let you set CPU and memory independently, and tune the DataFusion session parameters that control query partitioning, metadata reads, and output file layout.

Configure custom specifications for a DataFusion pipeline.

To configure a custom profile, open Build settings, select Custom under Profile size, and expand Configure custom specifications.

Custom profile settings

A custom profile exposes two kinds of settings:

• Resources: CPU cores and memory are the total compute available to your build. A standard profile pairs them at a set ratio; a custom profile lets you set each one on its own. Both are required.
• DataFusion session parameters: The remaining settings tune the query engine rather than the resources it runs on: how work is divided for parallel execution, how many files are read at once, and how rows are distributed across output files. Each setting is optional, and leaving one unset keeps the DataFusion default.

Custom profile use cases

A custom profile is useful when a fixed size does not fit the shape of your workload:

• Decouple memory from cores: A pipeline that needs a large working set but little parallelism can request 2 cores with 60 GB of memory, rather than moving up to a larger fixed profile to reach the memory it needs.
• Scale CPU past the largest fixed profile: A custom profile accepts up to 32 cores.
• Increase query concurrency. Raise Target partitions above the default of one partition per core.
• Control output file layout: Use Soft max rows per output file and Minimum parallel output files to influence how many output files a build writes and how large each one is.

Available specifications

• CPU Cores: Number of virtual CPU cores assigned to the executor.
• Memory: Amount of memory assigned to the executor.
• Metadata fetch concurrency: Number of files to read in parallel when inferring schema and statistics.
• Minimum parallel output files: The minimum number of output files written in parallel.
• Soft max rows per output file: A soft limit on the number of rows per output file. Each output file will contain roughly this many rows.
• Target partitions: Number of partitions for query execution. Increasing partitions can increase concurrency.

Standard Spark-backed pipelines and faster pipelines define compute profiles differently, so the custom specifications described above apply only to faster pipelines.

Learn more about compute profiles in Pipeline Builder.

Share your feedback

As we continue to add features to Pipeline Builder, we want to hear about your experiences and welcome your feedback. Share your thoughts with Palantir Support channels or our Developer Community using the pipeline-builder tag.

Workflow Lineage now includes a failure rate color mode

Workflow Lineage now includes a failure rate color mode that colors every action, function, and AIP Logic function node in your graph by the proportion of its runs that failed over a selected time window. This new feature shows where failures are concentrated across a given workflow, without needing to open each individual resource to check execution metrics.

Failure rate color mode is generally available on all enrollments starting the week of August 17, 2026.

A Workflow Lineage graph colored by failure rate color, with failed run counts annotated on affected nodes.

Color your graph by failure rate

Open the color legend in the Workflow Lineage toolbar and select Failure rate under the Health group. Every action, function, and AIP Logic function node in your graph is then colored by its failure rate over the selected time window:

• Less than 1% failure rate
• 1-5% failure rate
• 5-15% failure rate
• Greater than 15% failure rate
• No runs: The resource had no executions in the selected time window or is not supported.
• No permission: You are not a Viewer on the resource, so its metrics are not visible to you. Review the log permissions documentation for more information.

Nodes with at least one failed run are annotated with their run counts (for example, 12/430 runs failed), so you can distinguish a resource that failed 3 times out of 5 from one that failed 300 times out of 500. Node types that do not execute, such as object types, Workshop applications, and language models, are not colored by this mode.

Choose the time window you care about

Use the Date range selector in the color mode configuration to set the window over which the failure rate is calculated. The default is the past four hours, and you can select any window up to one day, either as a shortcut duration or as a specific date range. Coloring refreshes approximately once a minute while the graph is open in a visible tab. If you select a relative window, such as the past four hours, that window moves forward with each refresh. A fixed date range will remain where you set it.

The same color mode supports different questions depending on the window you choose. A four-hour window shows current failures, useful during an active investigation. A one-day window includes overnight scheduled activity, which is helpful when checking whether an intermittent failure from a previous day has recovered.

When to use failure rate coloring

Failure rate coloring is a graph-wide triage view; it shows which resources in a workflow are failing and how failures are distributed. You can then select a node to view its execution metrics, execution history, and logs to determine why the failures are happening.

Learn more about metrics and observability in Workflow Lineage.

Tell us what you think

As we continue to develop Workflow Lineage, we want to hear about your experiences and welcome your feedback. Share your thoughts with Palantir Support channels or our Developer Community using the workflow-lineage tag.

Features

AI Platform (AIP) / Language Model Service

Gemma 4 31B is now available in AIP through AWS Bedrock for eligible commercial enrollments with Bedrock enabled. Gemma 4 26B-A4B is coming soon.

AI Platform (AIP) / Language Model Service

Gemini 3.7 Flash is now available in AIP for non-georestricted, US georestricted, EU georestricted, IL2, IL4, and IL5 enrollments with Google Vertex AI enabled.

AI Platform (AIP) / Language Model Service

Grok 4.6 is now available in AIP on US georestricted and non-georestricted enrollments with xAI enabled.

Data connectivity & integration / Media sets

Working with media in functions is now generally available. Functions can accept media items as inputs, upload new media, attach media to Ontology objects through Ontology edits and function-backed actions, and read both the bytes and the metadata of any media item. These capabilities are available for TypeScript v2, Python, and TypeScript v1 functions.

A function reads media from media sets; this means that functions can use the documents, images, audio, and video that have already been ingested without having to export them somewhere else.

Data connectivity & integration / Pipeline Builder

Optimize your faster pipeline builds with custom compute profile configurations.

Ontology building / Workflow Lineage

Workflow Lineage now includes a failure rate color mode that colors every action, function, and AIP Logic function node in your graph by the proportion of its runs that failed over a selected time window. This new feature shows where failures are concentrated across a given workflow, without needing to open each individual resource to check execution metrics.

Features

Data connectivity & integration / Pipeline Builder

Pipeline Builder now allows users to edit the API name of object types directly within the object type editor. Previously, changing the API name of an object type owned by a pipeline required disowning the object type, modifying the API name in Ontology Manager, and then reclaiming it. The new field validates API names on submission, checking for duplicates within the current ontology and ensuring names conform to valid formatting requirements, including a maximum of 100 characters.

Use case development / Workshop

Workshop now includes a Discard unsaved changes option in the Save button's dropdown menu. The menu item appears only when there are unsaved changes to discard. Selecting this option opens a confirmation alert to prevent accidental loss of work, which mirrors the existing discard functionality available from the auto-save indicator in the top left of the page.

Use case development / Workshop

Workshop's Comments widget now exposes a Users tagged in comment option in the provided values dropdown menu for actions triggered after a comment is submitted. This string list contains only the user IDs of people @-mentioned in the comment, unlike the existing Users to notify value which includes all channel subscribers. By mapping an action's recipients parameter to this new value and disabling Send default notifications, builders can target only the users explicitly tagged in a comment. Duplicate mentions resolve to a single entry, and the list is empty when no users are mentioned.

Enhancements

Analytics / Notepad

Users may now embed images in envelope-secured Notepad documents on CBAC stacks. See Notepad envelope security documentation for more details.

Developer toolchain / Code Workspaces

Code Workspaces added two generally available compute-management capabilities: the compute selector now scales to a project's maximum queue CPU and memory instead of the previous 8 CPU / 64 GB limit, and resource administrators can now manage compute profiles in Control Panel by setting an enrollment-wide default profile and importing larger profiles into specific projects, while users in a managed enrollment select from a compute profiles dropdown limited to the profiles imported into their project.

Ontology building / Ontology Management

Ontology Manager now routes you to the entity's actual owning branch when you navigate to an object type, link type, action type, shared property, or interface that does not exist on your current branch. Previously, the switch prompt always suggested switching to the main branch, which could lead to a dead-end if the entity had not yet been merged. The prompt now displays the correct target branch name (e.g., "Switch to my-feature-branch").

Additionally, navigating to an entity in a different ontology while on a branch now correctly shows the switch ontology prompt instead of a generic "entity not found" error. An intermittent failure where discarding changes and switching branches could leave the ontology in a failed state has also been fixed.

Ontology building / Workflow Lineage

The Foundry Iceberg Maintenance page now displays a tag on each build in the history list identifying the maintenance operation that ran. Hovering over a tag highlights the corresponding task card in the right panel, making it easy to see which configured task produced a given build.

Use case development / Workshop

Workshop's event selector in the widget configuration sidebar now supports aliases when searching for events. The selector fuzzy-matches against both the event name and any configured aliases, so searching a common term like Reload surfaces the Refresh data in module event. Several commonly searched aliases have been added to existing events to improve discoverability.

Fixes

Data connectivity & integration / Pipeline Builder

Pipeline Builder now supports enabling the "Allow edits to objects of this type" toggle on object type outputs of Faster pipelines. Previously, this option was only available for other pipeline types. Users can now configure object type outputs in Faster pipelines to allow manual edits to the objects they produce.

AIP Analyst now supports new ways to capture reusable knowledge, analyze time series data, work with Ontology interfaces, and export results to Foundry resources.

Capture reusable knowledge with AIP skills and past analyses

Skills provide a first-class way to capture and share organizational knowledge. Users can configure which skills AIP Analyst has access to in the settings menu. AIP Analyst can load existing skills when they are relevant, create new skills from an analysis via the new export options, and modify existing skills. Skills are also supported as inputs to the AIP Analyst Workshop widget.

Additionally, the new Analysis lookup tool turns past analyses into a source of reusable knowledge. AIP Analyst can load a recent or favorite analysis by its RID directly into context, letting it understand how a related question was previously answered. The loaded analysis acts as a template: it shows what resources and tools were used, but carries no live results, so any relevant tool is rerun in the current analysis and always reflects current data.

Export conversations into Foundry resources

AIP Analyst now supports exporting conversations into a Notepad document, a Quiver analysis or dashboard, a Contour analysis, or an AIP skill. These tools are available in a redesigned Export button, which also includes the Print to PDF option. They are also available in-chat, by using a prompt such as “Export this conversation to a Notepad document”. These tools also support modifying existing resources, so you can ask for example “Update this notepad {notepadRid} with the results from this conversation”. To add these resource types to the conversation, select Add context, or drag and drop the resource into the chat window.

Analyze time series data

You can now import time series into your AIP Analyst chat in one of three ways:

• Add an object type with a time series property
• Add a time series property through the Add context button
• Add a time series sync using the Add context button

A full suite of time series transformations are available to manipulate the data, including filters, aggregates, mathematical operations, and event-based operations. You can export these time series charts into corresponding Quiver analyses and dashboards or included as Notepad images.

Work with Ontology interfaces

AIP Analyst now supports Ontology interfaces as a first class resource. You can import interfaces, search for them like object types, and use interface-scoped object sets in all of the regular object-set-supporting tools. Add them using the Add context button, or continue with your analyses as usual and they will be discovered by the object type search tool.

Alongside this change, AIP Analyst can now support multi-scope object sets. Unions, intersections, and differences can be created within AIP Analyst and consumed by most downstream tools, with warnings thrown if they are not. Both of these changes are compatible with Workshop.

We want to hear from you

We welcome your feedback on AIP Analyst, and your posts help the team prioritize what to build next. Share your thoughts with Palantir Support channels or our Developer Community using the aip-analyst tag.

Key features

AIP Evolve is now available in beta for enrollments with AIP enabled and access to AI FDE. AIP Evolve coordinates fleets of AI FDE agents to improve AI systems in AIP. Define a target, optimization goal, validation strategy, and operational constraints, then review the resulting proposal and agent activity before merging changes.

Early adopters of AIP Evolve have used it to autonomously cut AI costs, improve eval performance, and migrate workloads to open source models.

Learn more about AIP Evolve.

AIP Evolve supports iterative improvement workflows with the following features:

• Guided setup for model migration, cost reduction, latency reduction, evaluation score improvement, and custom goals.
• Flexible validation using selected test cases or existing evaluation suites.
• Configurable scoring criteria and acceptable output divergence.
• Agent constraints that control permitted change types and the maximum number of iterations.
• A proposal view containing proposed changes, validation results, output comparisons, supporting evidence, and confidence assessments.
• An interactive agent graph for monitoring progress and inspecting agent goals, insights, and artifacts.
• Integration with Global Branching for reviewing proposed changes before merging.
• The ability to resume an evolution in AI FDE with additional instructions.

Requirements

To use AIP Evolve:

• Enable AIP and ensure that you can access AI FDE.
• Install the AIP Evolve Marketplace product in an Ontology.
• Ensure that you have access to the target resource and any data or evaluation suites used for validation.

Getting started

The AIP Evolve setup screen, showing the "Review" stage with a fully specified evolution.

Open AIP Evolve and select New to create an evolution. Select the Foundry resource you want to evolve, then configure the following:

• Goal: Choose a predefined optimization goal or describe a custom objective.
• Validation strategy: Define the test data, scoring approach, and acceptable output divergence.
• Agent constraints: Select the types of changes agents may propose and set an iteration policy.

Review the generated prompt, then select Evolve. AIP Evolve opens AI FDE in a new tab and starts the evolution. You can also select Write custom prompt to provide your own instructions.

An example AIP Evolve proposal, where the system presents a model swap for cost reduction and the evidence that supports why the change is safe to make.

Open Evolutions to monitor active and completed evolutions. Use Proposal to review results and proposed changes, or Agent graph to inspect the agents involved in the evolution. When a proposal is ready, open it in Global Branching for final review and merging.

The AIP Evolve agent graph, showing each of the subagents that were spawned along the way to optimize the target AI component.

Your feedback matters

As we continue developing AIP Evolve, we welcome feedback about your experience. Share your thoughts through Palantir Support channels.

Applications

AI Platform (AIP) / AIP Evolve

AIP Evolve is now available in beta for enrollments with AIP enabled and access to AI FDE. AIP Evolve coordinates fleets of AI FDE agents to improve AI systems in AIP. Define a target, optimization goal, validation strategy, and operational constraints, then review the resulting proposal and agent activity before merging changes.

Early adopters of AIP Evolve have used it to autonomously cut AI costs, improve eval performance, and migrate workloads to open source models.

Learn more about AIP Evolve.

Features

AI Platform (AIP) / AIP Analyst

AIP Analyst now supports new ways to capture reusable knowledge, analyze time series data, work with Ontology interfaces, and export results to Foundry resources.

Features

Product delivery / DevOps

Marketplace now supports detaching outputs from an installation. Detached outputs are no longer marked as owned by Marketplace or tied to the installation. When an installation with detached outputs is edited or upgraded, Marketplace creates a new resource with a new identifier for each detached output. You can detach outputs from the Outputs view of a draft or from the installation details page.

Enhancements

Data connectivity & integration / Pipeline Builder

Pipeline Builder now includes an Explore data lineage action in the Files menu. This action opens the active pipeline branch in Data Lineage in a new tab, where users can inspect deployed inputs and outputs. The action is available when a saved pipeline has an active sandbox.

Fixes

Use case development / Slate

Slate now correctly handles rare timing issues where a function in progress could become stuck indefinitely, blocking all subsequent functions until the page was refreshed. This could occur when the function runner iframe reloaded during execution, causing pending functions to never complete. Functions now reliably continue executing across page navigation, View/Edit mode switches, and iframe reloads.

CoProcess user-defined functions (UDFs) are now supported in Pipeline Builder. A keyed CoProcess UDF allows users to define custom stateful logic for processing events from two different streams and merging them into a single output stream.

A keyed CoProcess UDF takes a separate stream on each of its left and right inputs.

Rows can be processed and emitted from either stream, which allows sophisticated real-time pipelines that go beyond standard join patterns. As with any stateful UDF, keyed CoProcess UDFs can be imported into a pipeline and used like any other transform.

Configure a keyed CoProcess UDF by mapping the columns of each input to the UDF.

What's new

With keyed CoProcess UDFs, you can:

• Maintain custom state across both input streams. State is shared between the two inputs rather than scoped to a single stream.
• Emit output rows from either stream independently. Neither stream has to wait on the other to produce output.
• Implement complex matching logic that standard joins cannot express.

Example use cases

• Flight tracking: Join a stream of live aircraft positions with flight plan schedule updates to detect route deviations in real time.
• Supply chain: Join shipment tracking events with inventory updates to suggest restocking before a warehouse runs out.
• Wildfire alerting: Join satellite thermal detections with ground-level weather data to escalate alerts when hotspots coincide with high-risk conditions.

Share your feedback

As we continue to add features to Pipeline Builder, we want to hear about your experiences and welcome your feedback. Share your thoughts with Palantir Support channels or our Developer Community using the pipeline-builder tag.

Features

Ontology building / Workflow Lineage

Workflow Lineage now supports Vertex resources. Both Vertex graphs and Vertex graph templates appear as nodes on the graph and can be found using search. You can navigate to these resources directly from Workflow Lineage.

Enhancements

AI Platform (AIP) / AI FDE

AI FDE now includes a button in the chat outline header that opens resources from your current conversation in a Workflow Lineage graph. This provides quick access to Workflow Lineage directly from the outline panel without needing to use the keyboard shortcut.

Ontology building / Ontology Management

Ontology Manager now displays link type constraints on interface pages in a single-sided format, showing only the target entity rather than repeating the current interface name on every row. Target names are now clickable links that navigate directly to the corresponding interface or object type page. This applies to the

Links

tab, the interface overview links section, and the interface extension preview. Previously, both sides of the constraint were displayed on each row and neither was clickable, requiring users to search for the target entity by name.

Security & governance / Checkpoints

The checkpoint record review interface now includes an interaction RID filter in the filter bar. Users can paste an interaction RID to filter records directly, with options to remove individual filters or clear all at once. This filter was previously only accessible via URL parameter.

Security & governance / Sensitive Data Scanner

Sensitive Data Scanner now allows you to exclude a resource directly from a recurring scan's results page. Each result row includes an action to navigate to the scan configuration's resource selection step, where the resource is pre-staged for exclusion. The exclusion is not applied automatically — you must review and save the updated configuration before it takes effect.

Use case development / Workshop

Workshop now resolves object set parameters passed to bidirectional iframe widgets more quickly. Updates to an object set in Workshop reach the embedded iframe more promptly without any change in behavior.

Fixes

Ontology building / Ontology Management

Ontology Manager no longer incorrectly blocks moving ontology resources into projects that have no maximum classification constraint. Previously, migrating an entity into such a project could produce a false constraint violation, preventing the migration from proceeding. Projects with explicit classification constraints continue to block real violations as expected, and metadata or query failures remain safely blocked.