Palantir has added experimental support in AIP for DeepSeek V4.1 Flash (1017), served through Fireworks. The change appears in an announcement on its Foundry documentation site dated November 3, 2026. Support covers non-georestricted commercial enrollments, and inference is hosted in the United States.
What changed
Palantir describes the model as open-weight and multimodal, designed for coding, reasoning and agentic workflows. In AIP, support includes text and image inputs, tool calling and a 1-million-token context window.
The status is experimental. Palantir said availability depends on the required releases reaching an enrollment and the relevant model settings being enabled. Enrollments that are georestricted or not commercial are not covered by the announcement.
How to get started
Palantir's announcements page lists three steps:
- Confirm that the enrollment administrator has enabled the DeepSeek model family through Fireworks and experimental models.
- Review token costs and pricing.
- See the complete list of models available in AIP.
The announcement does not give token prices, release numbers or a timeline for wider availability.
The same page carries other updates dated November 3. Palantir said "Web search is supported through Direct OpenAI and Amazon Bedrock on eligible enrollments" for supported OpenAI models in AIP. It also said "Faster pipelines are now the default in Pipeline Builder, with new support for Use LLM nodes."
Why it matters
Foundry News's view: this is a small entry, but it adds an open-weight option to the AIP model lineup. It also states where inference runs, a detail that may matter to buyers with data-residency requirements.
Tool calling plus a 1-million-token context window is the obvious draw. A team building an agent that works across a large codebase or document set could compare it with its current models on cost and output quality.
For now the support is experimental and gated by releases and administrator settings. We will watch for a change in that status and for any word on georestricted enrollments.
