Researchers from Snowflake and the University of Washington released MidTool on August 20, a 20.3B-token open corpus and pipeline for training LLMs on tool use between pretraining and task-specific fine-tuning. Mid-training Qwen3-4B-Base and Qwen3-8B-Base on MidTool-Mix produced models that outperform Qwen3's official instruction-tuned releases on MCP Universe at both parameter counts, before any task-specific post-training.
The central claim: knowledge for reliable tool use—when to call a function, how to extract arguments from noisy context, how to recover when a field is missing—lives in developer documentation, API specs, code repositories, and PDFs, not in curated agent trajectories. Post-training sees too little of this material too late. MidTool moves exposure earlier.
The pipeline builds MidTool-Mix from four sources: web pages, PDFs, code repos, and structured tool artifacts. Two synthesis branches convert raw material into training signal. Context-grounded trajectory augmentation takes documentation and code to produce supervision for recognizing tool boundaries, inferring parameters, and reconstructing workflow structure from messy, incomplete text. Native agentic trajectory synthesis builds executable multi-turn trajectories from live APIs and MCP skill definitions, with explicit validation of schema grounding, argument completeness, turn ordering, and tool-response consistency.
The corpus targets four atomic capabilities: recognizing tool affordances from natural language, grounding call arguments from long noisy context, composing multiple tools into sequential workflows, and recovering when required information is absent. These map directly to production failure modes—malformed JSON on nested schemas, hallucinated argument values, abandoned tasks when an API returns null. Standard SFT addresses the happy path; MidTool encodes error cases.
| Atomic Capability | What the Model Must Do | Production Failure Mode Addressed |
|---|---|---|
| Recognizing tool affordances from natural language | Identify when and which tool to invoke | Wrong or missed tool selection |
| Grounding call arguments from long noisy context | Extract correct parameter values from messy input | Hallucinated argument values |
| Composing multiple tools into sequential workflows | Chain tool calls in the right order | Malformed JSON on nested schemas; broken multi-step sequences |
| Recovering when required information is absent | Handle missing fields without abandoning the task | Abandoned tasks when an API returns null |
After mid-training, the team applies conventional post-training under both SFT and RL, evaluating on BFCL, τ²-Bench, and MCP Universe. MidTool-Mix improves results under both pipelines across all three benchmarks. The headline: 4B and 8B mid-trained models beat Qwen3's official instruction-tuned models on MCP Universe. The paper frames mid-training as compounding with, not substituting for, post-training—the mid-training stage installs a structural prior that improves subsequent SFT and RL efficiency.
A related mid-training paper (arxiv 2607.12463) tested a Python-only FIM corpus on coding agents and found it transferred gains to τ-bench and BFCL despite containing zero tool-use trajectories. The proposed mechanism: function calls and agent action-observation steps share a four-part structure. MidTool targets this isomorphism directly, explaining the 20.3B-token size versus the 2.6B-token coding corpus—tool-use knowledge is more heterogeneous than code structure.
Corpus, models, and pipeline are released at hf.co/collections/MidTool/midtool-release. For architects deciding whether to build tool-calling into a base model before task-specific fine-tuning, MidTool provides a reproducible recipe with public weights at two parameter scales that benchmark against an official baseline.