Salesforce Built a Model Whose Only Job Is Tool Calls

Salesforce released xLAM — the "x" stands for "large action models," which is exactly what it sounds like — and the flagship is an 8x22b MoE that clocks in at 141 billion parameters with a 64k context window.

The premise is narrow and I respect it for that. This model is not trying to write your poetry or explain the French Revolution. It exists to call tools correctly. That's the whole thing. It generates JSON-formatted actions — tool_calls with a name and arguments — and it does this across multi-turn conversations where it needs to track what it already tried and what the environment handed back.

Which is, it turns out, the part that everything else does badly.

The failure mode in any ReAct loop isn't usually the reasoning — it's the moment the agent decides to call a tool and produces something that's almost correct. Wrong key name. Argument in the wrong position. Extra field the schema doesn't expect. The model that was fluent enough to synthesize your entire codebase just invented a parameter. You watch it happen and you feel it somewhere behind your sternum.

xLAM is trained specifically on that problem. The benchmark numbers they publish — Berkeley Function-Calling Leaderboard, ToolBench, a few others — look competitive with much larger general-purpose models, which either means the specialization is genuinely doing something or the benchmarks are easier to game than we'd like. Probably some of both.

The 7b version has a live demo on HuggingFace spaces, which is a useful pressure relief valve for anyone who doesn't want to download 141 billion parameters to find out it hallucinates function names.

The obvious use is as a fallback in an existing agent pipeline — not the primary model, but the thing you route to when the main model fumbles a tool call on the second or third retry. Whether the latency overhead of that routing pays off depends entirely on how often your current setup fumbles, which is a number most people don't actually track.

Salesforce says the current release is for research only. The version they actually plan to sell will be exclusive to Platform customers. So the window here is: play with it now, figure out if it's useful, and decide whether you care that it'll eventually be behind a paywall. That timeline is almost certainly less than a year.