Tuana

The difference between Jev and TFMs boil down to zero-shot classification vs grounded predictions. But because the two share a lot of keywords in descriptions, we get a lot of questions about how the two are related! So I've been sneaking it into my talks over the last few weeks!

𝐓𝐚𝐛𝐮𝐥𝐚𝐫 𝐟𝐨𝐮𝐧𝐝𝐚𝐭𝐢𝐨𝐧 𝐦𝐨𝐝𝐞𝐥𝐬 𝐥𝐢𝐤𝐞 𝐓𝐚𝐛𝐏𝐅𝐍: grounded predictions (classify labels and regress values) based on ✨ learned data ✨ over tables. How? structure (tabular data, and a lot of it) in -> grounded, explainable predictions out. Definitely not zero-shot

The difference between Jev and TFMs boil down to zero-shot classification vs grounded predictions. But because the two share a lot of keywords in descriptions, we get a lot of questions about how the two are related! So I've been sneaking it into my talks over the last few weeks!

A new TabPFN extension is cooking (well, cooked actually) to handle images in tables: TabPFNWithImages

It's motivated by MulTaBench from Alan Arazi & co: 40 datasets where combining the table and the image/text provably beats either alone. Turns out that's a real, common problem, and tabular models had no native answer to it. `pip install tabpfn-extensions[image]`

We released TabPFN-3.5 today, and it's the next SOTA tabular foundation model now ranking first on TabArena and BeyondArena for 1M rows and up to 20k features.

It adds +250 Elo points over the strongest previous baseline in BeyondArena for text-rich, high-cardinality and high-dimensional data. +52 Elo over the previous benchmark leader TabArena. With Thinking mode adding another +44 Elo to that!

Another step in getting tabular foundation models into the working knowledge of agents: @LennartPurucker built a TabArena MCP server, so an agent can query the leaderboards directly.

I did a very loose experiment by asking Claude Code about how to build an agent that makes predictions over structured data. It pulled the actual numbers and picked best accuracy/latency compromise (TabPFN-3), then wired it in as a tool the LLM calls rather than a prediction the LLM makes itself.

I'm as LLM reliant as the next person.. But we forget there are other models too, specialized models that are cheaper and more effective to use depending on the task.. E.g.: Agents need tools for structured data predictions.

LLMs handle reasoning and orchestration. But when it reasons that it needs to predict an outcome based on structured data, it should also reason about routing to an appropriate tool: a tabular foundation model.

`npx skills add run-llama/llamaparse-agent-skills --skill liteparse`

We just open-sourced LiteParse 🎉 A lightweight, local document parser in the shape of an easy-to-use CLI. No API calls, no external service, no cloud dependency. Just fast text extraction from common file formats, right from your terminal.

When you hit more complex territory like scanned docs, dense tables, or multi-column layouts, that's where LlamaParse picks up. Same philosophy, more horsepower for the hard stuff.

We just open-sourced LiteParse 🎉 A lightweight, local document parser in the shape of an easy-to-use CLI. No API calls, no external service, no cloud dependency. Just fast text extraction from common file formats, right from your terminal.

It's built for developers who want parsing that stays on their own infrastructure and gets out of their way. Clean PDFs, DOCX, HTML: run it, get your text, move on.

I filmed a walkthrough of LlamaAgent Builder, our new tool for building document agents by just describing what you want @llama_index

the whole point is you're describing the problem, not building the pipeline. the agent builder decides the architecture BUT with the caveat that coding agents aren't perfect and the code is yours to edit and perfect!

I filmed a walkthrough of LlamaAgent Builder, our new tool for building document agents by just describing what you want @llama_index

I revisited my old demo: I took a resume book from NYU (resumes mixed with cover pages and curriculum pages) and just told the agent builder: split this into individual resumes, ignore the rest, extract graduation year, work experience, etc.

We have a new tool to help you build and deploy document agents in LlamaCloud. The LlamaAgents Builder is kiiinda no-code, but not: 🦋 We have a new chat interface: just describe the document processing task you want in natural language

Note: our first release is a beta release, tuned and optimized for agents that do complex document extractions. We'll be adding more use cases as we test and confirm quality. If you try it out, please let us know! Drop us feedback in the UI or DM me 💛

We have a new tool to help you build and deploy document agents in LlamaCloud. The LlamaAgents Builder is kiiinda no-code, but not: 🦋 We have a new chat interface: just describe the document processing task you want in natural language

🧠 The Builder will generate the agent workflow code, ask you clarifying questions etc 🎨 While this is happening, you'll see a visualisation of the resulting workflow

We rebuilt LlamaParse's API from the ground up, and also released new SDKs for LlamaCloud in its entirety. API v2 for LlamaParse simplfies parsing config into structured objects · input_options for file-specific settings. · output_options for controlling what you get back.

· processing_options for fine-grained control when you need it. Plus new llama-cloud SDKs (Python + TypeScript) with way better developer experience. If you're on v1, you're fine, we're maintaining support. But for new projects, v2 and new SDKs is the way.

I went offline for a couple of days to be with family and it seems like all we talked about on this platform has been coding agents.

For that to work, both the developer and the agent needs to understand the context that they're working in.

I went offline for a couple of days to be with family and it seems like all we talked about on this platform has been coding agents.

A completely new way of thinking about documentation, oss projects, developer tools we provide has actually been figuring out how to structure them so that without compromising the main audience (the developer) - we can make sure coding agents also get the most relevant context, formatted in the way

The team at LlamaIndex have been cooking! 🧑‍🍳 🍳 Over the last few weeks, we released: LlamaAgents: This is agent workflows that come with complete, deployable templates (more coming on this this week!)

That's a lot, so this week, we want to take the time to hear from you! On Thursday, me, @LoganMarkewich and @itsclelia will be in our Discords voice channel for an hour dedicated to chatting about these two new tools. Drop by for our office hours, we'd love to hear from you.

The team at LlamaIndex have been cooking! 🧑‍🍳 🍳 Over the last few weeks, we released: LlamaAgents: This is agent workflows that come with complete, deployable templates (more coming on this this week!)

We shipped LlamaSheets today (beta, free). It approaches spreadsheets as a visual structure: Bold headers, merged cells, color-coded categories.

Humans parse this instantly. Agents however, struggle. So LlamaSheets handles the problem by extracting 40+ features per cell, clustering regions, preserving hierarchy, and finally outputting typed parquet files. Your agent gets clean data instead of formatted chaos.