2.25.6

What's Changed

• Fix thread safety issues in MLX concurrent inference (Samplers) by @sxy-trans-n in #351
• Port of Ernie4 5 by @smdesai in #348
• Add SmolLM3 by @johnmai-dev in #349
• Add LFM2 by @johnmai-dev in #354
• Add Baichuan M1 by @johnmai-dev in #355
• Add Exaone4 by @johnmai-dev in #357
• feat: implement gemma3n text model in MLXLLM by @xlab in #346
• Add Deepseek V3 model by @danielnugraha in #249
• Remove duplicate keys by @tattn in #363
• Fix LFM2 by @DePasqualeOrg in #369
• fix: DeepSeek V3 configuration parsing and RoPE weight loading by @sxy-trans-n in #368
• fix: handle quantized tie-embedding for Gemma3Text by @jiwoong-choi in #372
• Add GPT OSS by @johnmai-dev in #371
• swift-format by @davidkoski in #376

New Contributors

• @sxy-trans-n made their first contribution in #351
• @xlab made their first contribution in #346
• @danielnugraha made their first contribution in #249
• @tattn made their first contribution in #363
• @jiwoong-choi made their first contribution in #372

Full Changelog: 2.25.5...2.25.6