xLAM-2

What it is

Salesforce describes large action models as the brains of AI agents: the part that turns an intent into executable actions ^[1]. xLAM-2 is the company's family of such models, spanning 1 billion to 70 billion parameters and built on Llama and Qwen base models ^[1][2]. The range of sizes matters in practice, since the same approach is meant to be usable both where compute is tight and where a large model can be served.

The design target is not the single function call that early tool-use demonstrations relied on. The models are tuned for multi-turn tool use rather than single-call function calling ^[1]. The difference is practical: a single call takes one request and returns one tool invocation, while multi-turn use means the model calls a tool, reads the result, decides what to do next, and repeats, sometimes over many steps and several different tools. That loop is what an agent actually runs on, and it is harder to learn because the model has to keep track of state across turns.

Where the training data comes from

The data side is described in a paper whose title spells out the pipeline: APIGen-MT, an agentic pipeline for multi-turn data generation via simulated agent-human interplay ^[2]. It runs in two phases. The first writes task blueprints, structured descriptions of what a task should involve. The second simulates an agent and a human interacting, and the resulting trajectories become training data ^[2].

Between the two phases sits a filter: a committee of LLM reviewers checks the blueprints before any trajectories are generated ^[2]. Putting the check at that point is an ordering that saves work. Generating multi-turn interaction is the expensive part, so screening blueprints first keeps bad task designs from consuming the budget.

Salesforce also published the APIGen-MT-5k and xlam-function-calling-60k datasets alongside the models ^[1], which makes the data work inspectable and reusable rather than hidden inside the release.

Results

Salesforce reports that the models beat GPT-4o and Claude 3.5 on BFCL and tau-bench ^[1]. The model card dates that benchmark comparison to 2 April 2025 ^[1]. These are comparative claims made by the releasing organisation rather than an independent evaluation, and they concern two tool-use benchmarks rather than a general capability ranking; readers who need absolute numbers should consult the source. What the claim does establish is the comparison class: xLAM-2 is being positioned against general-purpose frontier chat models, not only against other specialised action models.

Packaging for agents

A model that is good at tool use still has to be wired into a loop, and the release reflects that. The chat template and the vLLM integration were reworked for agent deployment ^[1]. The chat template is the piece that formats a conversation into the exact prompt a model expects, and an agent loop re-sends that prompt at every step with tool outputs appended; the vLLM integration is the serving path that answers those prompts. Both are the kind of plumbing that decides whether a model can be dropped into an agent framework without custom glue.

Licence and scope

The models are released under CC-BY-NC 4.0 and the release is labelled research-only ^[1]. The non-commercial term means the weights can be studied and used in research, but not put into commercial products as they stand.

Where it sits

xLAM-2 belongs to a shift in what a model's output is for: from producing text to producing actions. By framing the model as the brains of an agent, Salesforce treats action-taking as a capability with its own data pipeline, benchmarks and serving requirements — a generator that writes blueprints and then simulates interaction, a reviewer committee that screens the blueprints, tool-use leaderboards, and a chat template built for agent loops ^[1][2].