v0.28.3

What's Changed

• Removing the deprecated wandb params by @Goekdeniz-Guelmez in #524
• Memory efficient ssm by @awni in #525
• Fix lora MoEs by @awni in #522
• Fix bailing moe by @awni in #521
• GPT2 Batching Fix by @shepardxia in #529
• Remove act loss and add temp in DWQ by @awni in #500
• Cleanup and simplify model I/O by @awni in #532
• Fix: Add future annotations import to qwen3_next.py for Python 3.9 compatibility by @mzau in #533
• Add lfm2 moe by @Blaizzy in #537
• Fix example command to quantize a model using GPTQ by @felladrin in #539
• minor typing issues by @mercush in #540
• Fix cuda install by @awni in #542
• Add Qwen3-VL language model implementation by @vincentamato in #547
• Fix mask for batched SSM by @awni in #546
• Added gradient accumulation to training loop by @dotvignesh in #511
• Support data parallel eval for generation tasks by @awni in #549
• Optimize Bailing MoE by @kernelpool in #550
• Adding jamba by @Goekdeniz-Guelmez in #544
• Add Qwen3-VL (Dense) language model implementation by @vincentamato in #553
• LLM Benchmarks by @awni in #552
• Add support for nanochat by @dnakov in #554
• version by @awni in #558

New Contributors

• @shepardxia made their first contribution in #529
• @mzau made their first contribution in #533
• @felladrin made their first contribution in #539
• @mercush made their first contribution in #540
• @dotvignesh made their first contribution in #511

Full Changelog: v0.28.2...v0.28.3

v0.28.2

What's Changed

• fix bailing moe by @awni in #514
• Fix batching for models with nested cache structures by @kernelpool in #510
• Fix: Correct weight masking for zero-computation experts in LongCat Flash MoE by @kernelpool in #508
• Simplify to_lora to not hardcode model types by @awni in #515
• Add Olmo3 by @Goekdeniz-Guelmez in #445
• Make mixed quantization affect attention in DeepSeek V3, others by @n8sh1 in #506
• Add Apriel 1.5 by @ivanfioravanti in #520
• feat: Refactor granitemoehybrid to support dense and non-hybrid variants by @gabe-l-hart in #518

New Contributors

• @kernelpool made their first contribution in #510

Full Changelog: v0.28.1...v0.28.2

v0.28.1

What's Changed

• Fix quant predicate by @awni in #485
• Fix passing argument model_config in utils.load() by @ariahw in #494
• Fix KV cache quantization for hybrid models by @awni in #495
• fix for LFM2 by @awni in #493
• fix loading for qwen2 VL by @awni in #491
• Enable training for qwen3 next by @awni in #496
• Add batch support for sliding window cache by @awni in #487
• qwen3 next batching by @awni in #478
• Add Falcon H1 by @Blaizzy in #231
• Fix RotatingKVCache update by @awni in #503
• Add Code World Model support by @dnakov in #505
• Use depends in pipeline parallel by @awni in #483

New Contributors

• @ariahw made their first contribution in #494

Full Changelog: v0.28.0...v0.28.1

v0.28.0

What's Changed

• Allow fp8 by @awni in #431
• Avoid cache-trimming crash in server for longcat chat and baichuan_m1 by @n8sh1 in #434
• Fix hunyuan v1 dense by @awni in #440
• Changes needed to facilitate batching by @awni in #430
• remove manual conv class in mamba1 by @Goekdeniz-Guelmez in #436
• adding Kwai-Klear/Klear-46B-A2.5B-Instruct by @Goekdeniz-Guelmez in #437
• Add lille 130m by @Goekdeniz-Guelmez in #429
• model: GraniteMoeHybrid by @gabe-l-hart in #442
• fix server paths by @awni in #448
• sdpa with sinks by @awni in #418
• fix(quantization): Parameterize hardcoded group_size in mixed_quant_predicate_builder by @squaredice in #449
• Adding Ling Mini by @Goekdeniz-Guelmez in #450
• Adding Qwen3 Next by @Goekdeniz-Guelmez in #441
• Faster ssm by @awni in #451
• Update bitnet, nemotron h to use build in relu2 from MLX by @Goekdeniz-Guelmez in #446
• fix qwen3 next by @Goekdeniz-Guelmez in #453
• Adding GLM by @Goekdeniz-Guelmez in #457
• Add an introduction to the default LLM in README.md by @aopstudio in #461
• Fix TypeError: Model.__call__() got an unexpected keyword argument 'mask' for qwen2_vl, mistral3 by @neilmehta24 in #464
• Add groups to ssm kernel and update more models by @awni in #456
• Fix gemma3 window mask by @awni in #465
• Batch generation by @awni in #443
• Batch support for mamba-style models by @awni in #468
• fix: handle cache offset safely for mamba error by @ivanfioravanti in #472
• Adds LLaMA 4 text model implementation in MLX by @robbiemu in #469
• Allow sampler to work with batched_generate by @N8python in #473
• Adding support for mamba2 by @Goekdeniz-Guelmez in #392
• Fix llama4 text and make trainable by @Goekdeniz-Guelmez in #474
• Extends quantization predicate with config by @robbiemu in #476
• Gated-Delta Fused Kernel (Qwen3Next) by @ivanfioravanti in #454

New Contributors

• @gabe-l-hart made their first contribution in #442
• @squaredice made their first contribution in #449
• @aopstudio made their first contribution in #461
• @robbiemu made their first contribution in #469

Full Changelog: v0.27.1...v0.28.0

v0.27.1

What's Changed

• Fix mlx_lm.perplexity seed w/ np.random.seed to ensure determinism across runs by @N8python in #415
• server allow specifying seed by @n8sh1 in #414
• Adding ibm Granite MoE by @Goekdeniz-Guelmez in #413
• had to add self.args and self.model_type into the model class for mlx-lm-lore by @Goekdeniz-Guelmez in #422
• add Apertus from Swiss AI by @Goekdeniz-Guelmez in #421
• Add nemotron h by @Goekdeniz-Guelmez in #407
• Adding longcat flash by @Goekdeniz-Guelmez in #423
• Fix Nemotron H loading error by @Goekdeniz-Guelmez in #426

New Contributors

• @n8sh1 made their first contribution in #414

Full Changelog: v0.27.0...v0.27.1

v0.27.0

What's Changed

• Add mlx_lm.perplexity by @N8python in #397
• Benchmark script by @awni in #396
• Don't reload default model by @awni in #400
• only apply lm_head to the last token by @awni in #406
• Fix prompt cache corruption when generation is interrupted by @dojoteef in #405
• support mxfp4 by @awni in #385
• version by @awni in #410

Full Changelog: v0.26.4...v0.27.0

v0.26.4

What's Changed

• Revert symmetric kl by @awni in #359
• Adding bailing_moe (ling-lite, -plus, -coder) by @Goekdeniz-Guelmez in #369
• Add SwanLab experiment tracking support for MLX by @ShaohonChen in #317
• Fix gpt-oss lora nan by @awni in #370
• Properly tie embeddings and lm head for gemma3 by @awni in #373
• Fix distributed evaluate by @angeloskath in #368
• Add SSE keepalive to stop client disconnects during prompt processing by @dysangel in #362
• Add LFM2-VL model implementation by @christian-lms in #378
• Adding trust_remote_code=True for training by @Goekdeniz-Guelmez in #383
• remove comma and add muon by @Goekdeniz-Guelmez in #381
• add into the lora to layer utils by @Goekdeniz-Guelmez in #382
• Make KL and JS metal kernels only if metal is available by @vsabolcec in #387
• fix sampling with small top k by @awni in #388
• fix window attention mask by @awni in #390
• Add support for ByteDance Seed-OSS-36B-Instruct model by @dnakov in #391
• Add Qwen2-VL model implementation by @vincentamato in #384

New Contributors

• @ShaohonChen made their first contribution in #317
• @dysangel made their first contribution in #362
• @dnakov made their first contribution in #391
• @vincentamato made their first contribution in #384

Full Changelog: v0.26.3...v0.26.4

v0.26.3

What's Changed

• Resolve streaming last token error and correct total token usage by @zenyr in #342
• Fix NameError in loglikelihood_rolling method by @snellingio in #339
• fix error on unsupported response type in server by @emmanuel-ferdman in #344
• Add --trust-remote-code cli option by @dojoteef in #319
• Add validation set for DWQ by @awni in #343
• feat: add --confirm-run-unsafe-code CLI option to allow execution of untrusted code by @ivanfioravanti in #348
• Allow per model quant config by @awni in #349
• Add gpt_oss model by @christian-lms in #354
• Jensen-Shannon divergence loss kernel by @vsabolcec in #352
• Route the gpt_oss to fused sdpa by @angeloskath in #356
• Hunyuan V1 Dense model support by @ivanfioravanti in #351
• Add Additional Features of GPT-OSS Model : Lora, Alternating attention, MoE Support by @Shashikant86 in #357

New Contributors

• @zenyr made their first contribution in #342
• @snellingio made their first contribution in #339
• @emmanuel-ferdman made their first contribution in #344
• @dojoteef made their first contribution in #319
• @vsabolcec made their first contribution in #352
• @Shashikant86 made their first contribution in #357

Full Changelog: v0.26.2...v0.26.3

v0.26.2

What's Changed

• chore: fix gemma3n intermediate_size config by @mzbac in #332
• Add GLM4.5 by @awni in #333
• Changed GLM-4 MoE support for DWQ quantization by @ivanfioravanti in #336
• Add system prompt to chat script by @jussikuosa in #334
• Initialize empty cache when cache=None for Gemma‑3n models by @brchristian in #323

New Contributors

• @mzbac made their first contribution in #332
• @jussikuosa made their first contribution in #334
• @brchristian made their first contribution in #323

Full Changelog: v0.26.1...v0.26.2

v0.26.1

What's Changed

• Add dsv3 for lora by @awni in #284
• GPTQ quantization by @awni in #279
• Fix MoE fine tuning by @awni in #288
• fix hunyuan by @awni in #286
• Allow trust_remote_code in convert.py by @christian-lms in #289
• Type Signature Fixes by @MattBeton in #290
• Fix gemma3n config load bug by @neilmehta24 in #292
• kimi k2 by @awni in #293
• Add LFM2 by @Blaizzy in #291
• feat: DWQ for Hunyuan-A13B-Instruct and trust_remote_code argument by @ivanfioravanti in #303
• fix: update import for huggingface model in evaluate.py by @ivanfioravanti in #275
• Fix ddp workers loading the same data by @angeloskath in #294
• Fix server finish reason by @awni in #307
• Add support for SGD & Adafactor by @N8python in #306
• Allow empty prompt with input_embeddings by @will-lms in #308
• fix naive detokenizer by @awni in #312
• add exaone4 by @awni in #310
• add v1/models/repo_id by @awni in #313
• Update W&B logging crash in MLX-LM-LORA by @Goekdeniz-Guelmez in #316
• Fix DSV3 training by @awni in #324
• Lora works with cuda backend by @awni in #330
• Adding Muon Optimizer by @Goekdeniz-Guelmez in #325

New Contributors

• @christian-lms made their first contribution in #289
• @MattBeton made their first contribution in #290
• @N8python made their first contribution in #306

Full Changelog: v0.26.0...v0.26.1