Run Hermes Free on Gemma 4 After Google's Tool-Calling Fix

Key Takeaways

• The 2026-07-15 refresh is weights-only — same version number, new files on Hugging Face. If you pulled Gemma 4 before that date you do not have the fix; re-pull.
• Tool calling was the headline fix. Google's own testing is cited as showing one use case gaining over 10% accuracy. That's the difference between an agent you can leave running and one you have to babysit.
• Speed: prompt processing up to 70% faster on newer GPUs, and up to 31% faster time-to-first-token.
• The crash cause was an attention-size mismatch. 26 of Gemma 4's layers use one attention size that standard acceleration handles fine; 4 layers use a larger size those same kernels couldn't handle. That mismatch drove the local crashes and stalls — the refresh closes the gap.
• Better image handling — a raised detail setting supports images over 2.5× larger, which matters if you're OCR-ing documents, screenshots or handwritten notes.
• Wiring it into Hermes: open the Hermes dashboard → Manage → Add profiles → pick your model. Install Ollama and Gemma 4 appears ready to use; alternatively pull the weights from Hugging Face and serve them through LM Studio.
• The cost argument: an agent loop that thinks, acts, checks its own work and loops again burns tokens on every pass. Running that loop locally costs nothing per iteration, can't be rate-limited mid-task, and keeps working if your connection drops.
• The honest caveat, stated in the video: Gemma 4 does not match a frontier model like Fable 5 on raw intelligence. The recommended shape is a split — a stronger model plans and outlines, Gemma 4 grinds through the repetitive execution for free.