Google Gemma 4 发布并支持在 Unsloth 中微调

Google DeepMind releases Gemma 4 today, a new family of open models, including E2B, E4B, 26B-A4B, and 31B. These multimodal, hybrid-thinking models support 140+ languages, up to 256K context, and come in both dense and MoE variants. Gemma 4 can now run locally and be fine-tuned in Unsloth Studio.

Thank you Google for providing Unsloth with day zero access. For more details on Gemma 4 and usage, read our blog / guide:

Gemma 4 Blog

Since we released Unsloth Studio 2 weeks ago, we’ve made many improvements. Most notably:

• 50-80% more accurate Tool calling: better parsing, better accuracy, faster execution, no raw tool markup in chat, plus a new Tool Outputs panel and timers.

• Pre-compiled llama.cpp binaries for ~1min installs and -50% less size

• Auto-detection of existing models / GGUFs

• 20–30% faster inference, now matches llama-server / llama.cpp speeds.

• New Desktop app shortcuts

• Data Recipes now supports macOS, CPU and multi-file uploads.

Unsloth Studio

Unsloth Studio now also installs in one command:

macOS, Linux, WSL:

curl -fsSL https://unsloth.ai/install.sh | sh

Windows:

irm https://unsloth.ai/install.ps1 | iex

A new way to use Unsloth is also coming soon…

No posts