MidTool Moves Tool Use into Mid-Training, Raising Qwen3-4B’s BFCLv3 Overall Score by 10.5 Points

Tool use is typically addressed during supervised fine-tuning or reinforcement learning. Researchers from Snowflake, the University of Washington, and other institutions introduced MidTool, arguing that this capability should be moved earlier, into a mid-training stage between foundation pretraining and post-training. MidTool-Mix contains 20.3 billion tokens across 11.22 million samples, with web pages, PDFs, code, and native agent trajectories accounting for 42%, 23%, 26%, and 9%, respectively.

The data pipeline does not simply feed documents directly to the model. The researchers first apply fastText, hashing, and MinHash LSH for filtering and deduplication, then extract tool purposes, schemas, parameters, and workflows from the documents. A separate branch uses REST APIs, MCP skills, and actual rollouts to synthesize executable, multi-turn trajectories. Tool calls are also checked with deterministic rules for turn ordering, required parameters, and response consistency.

The team mid-trained Qwen3-4B-Base and Qwen3-8B-Base separately on 32 H200 GPUs, then performed SFT using the same 100,000-example TOUCAN dataset. Some variants additionally ran GRPO across 526 synthetic environments. For the 4B model, adding only MidTool-Mix and SFT raised the BFCLv3 overall score from the SFT baseline of 39.73% to 50.25%; adding RL increased it further to 54.18%. The best 8B variant improved from an SFT baseline of 47.62% to 55.12%. On MCP-Universe, which more closely reflects real-world MCP servers, the best 8B variant achieved an overall score of 25.16, compared with the baseline’s 15.18.

The results support the idea that tool capabilities do not need to be injected entirely during post-training, but they do not yet represent a breakthrough toward general-purpose agents. Pass rates for MCP-Universe’s multi-server and web-search tasks remain at zero, while performance on telecommunications tasks is also weak. Several datasets were synthesized using GPT-5-series and Qwen teacher models, so costs, teacher bias, and contamination audits still require independent reproduction. Engineering teams should next examine whether the public data can be downloaded in full and whether the same gains persist across model families, tool formats, and real-world enterprise APIs.