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.

MidTool corpus pipeline: four data sources feed two synthesis branches to produce the 20.3 B-token MidTool-Mix.
FIG. 02 MidTool corpus pipeline: four data sources feed two synthesis branches to produce the 20.3 B-token MidTool-Mix. — MidTool paper (arxiv 2608.20314)

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 CapabilityWhat the Model Must DoProduction Failure Mode Addressed
Recognizing tool affordances from natural languageIdentify when and which tool to invokeWrong or missed tool selection
Grounding call arguments from long noisy contextExtract correct parameter values from messy inputHallucinated argument values
Composing multiple tools into sequential workflowsChain tool calls in the right orderMalformed JSON on nested schemas; broken multi-step sequences
Recovering when required information is absentHandle missing fields without abandoning the taskAbandoned tasks when an API returns null
FIG. 03 MidTool's four atomic tool-use capabilities and the production failure modes each one targets. — MidTool paper (arxiv 2608.20314)

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.

Three-stage training pipeline: mid-training on MidTool-Mix is inserted between base pretraining and task-specific post-training.
FIG. 04 Three-stage training pipeline: mid-training on MidTool-Mix is inserted between base pretraining and task-specific post-training. — MidTool paper (arxiv 2608.20314)

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.

MidTool-Mix (20.3 B tokens) is ~7.8× larger than the comparable Python FIM coding corpus (2.6 B tokens), reflecting the greater heterogeneity of tool-use knowledge.
FIG. 05 MidTool-Mix (20.3 B tokens) is ~7.8× larger than the comparable Python FIM coding corpus (2.6 B tokens), reflecting the greater heterogeneity of tool-use knowledge. — MidTool paper (arxiv 2608.20314); related coding corpus paper (arxiv 2607.12463)

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.