wtf is best-local-model_agentic-workflows_2026?

This repository is not a piece of software to run, but an interactive comparison report that helps someone pick which local language model to use as the backend for agentic workflows, meaning multi step tool use, function calling, and autonomous planning as done by frameworks like Hermes Agent. It targets people running large local models on high memory hardware, such as a single DGX Spark machine or similar rigs with 96 to 128 gigabytes of memory. The benchmark data behind the report comes from a separate project called tool-eval-bench, which ran 84 test scenarios across 16 categories, with up to 8 trials per model, scoring each attempt as pass, partial, or fail. The README's quick answer names one model, Qwen 3.6 35B A3B UD Q8_K_XL, as the recommended default backend, citing a 91.0 mean score and a 100 percent pass rate across every trial in core categories like multi-step chains, error recovery, tool selection, and structured output. A full ranking table lists seven models total, each with a score and a short note on its strengths or weaknesses, plus a tier recommendations section suggesting different models for default production use, safety critical workloads, and low latency needs, while also naming two models the author advises against deploying unattended due to reliability or safety issues found during testing. Viewing the actual report requires no build step or server. It is delivered as two static HTML files, a dark themed primary version and a light themed alternate, that can be opened directly in a browser after cloning the repository. A methodology section explains the scoring approach, including two specific reliability metrics: one measuring whether a model solves a task at least once across trials, and another measuring whether it solves the task every single time. The README and reports are released under the MIT license, while the underlying benchmark data belongs to the respective model evaluators.