Microsoft's Phi-4 now in Unsloth!

Hope you all had an amazing New Year! Here's our latest updates:

---

✨ Microsoft Phi-4

• Phi-4, Microsoft's new 14B model that performs on par with OpenAI's GPT-4o-mini is now in Unsloth!

• We fixed 4 bugs in Phi-4, greatly increasing the model’s accuracy.

• Unsloth Phi-4 finetuning enables >128K context lengths

---

🦙 Llama 3.3 + Ultra long context

• Llama 3.3 is Meta’s new 70B model that performs on par with Llama 405B.

• We worked with Apple to integrate CCE, making Unsloth now support 13x longer context lengths.

---

🦥 Unsloth Dynamic 4-bit

• We’re also excited to introduce Unsloth Dynamic 4-bit Quantization!

• Unlike naive quantization, which reduces accuracy, our method selectively avoids quantizing certain parameters, achieving significantly higher accuracy, while only using <10% more VRAM than BitsandBytes 4-bit.