🧰 Transformers

The items that name it 31

• Welcome Le Chonk: Mistral Large 4 on Hugging Face
  articleas frameworks and tools2026-10-29

  Hugging Face library with day-0 support for Mistral Large 4 via AutoModelForMultimodalLM and AutoProcessor; supports distributed inference with tensor and expert parallelism.

• Build specialized multimodal models with LFM3
  articleas classification.technologies2026-10-27
• Folding Attention Biases into Keys Makes Compacted KV Caches Run on Standard Kernels
  paperas classification.technologies2026-10-27
• Syb (@DegenOfWallSt) on X
  threadas frameworks and tools2026-10-25

  Library that loads Nanbeige4.3-33B with trust_remote_code.

• Does Your Model Run on AWS Trainium? Look It Up
  articleas frameworks and tools2026-10-23

  Library whose models are tested; loads models with from_pretrained and runs its own modeling code on Trainium.

• Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster
  articleas frameworks and tools2026-10-21

  Native support for all three Granite Speech models.

• Hugging Face ❤️ TorchTPU: TPUs, Natively
  articleas frameworks and tools2026-10-20

  Hugging Face model definition library (v5). Loads models on TPU with device_map="tpu".

• MPS in PyTorch 2.15
  articleas frameworks and tools2026-10-20

  Hugging Face library for running models on a Mac with PyTorch.

• TorchNeuron: running HF models on AWS Trainium is easier than ever
  articleas frameworks and tools2026-10-19

  Hugging Face library that runs text generation and training on Neuron via device_map=neuron and DistributedConfig.

• DeepSeek-v4.1-Flash-1017
  repoas frameworks and tools2026-10-17

  Library used to load the model via pipeline (image-text-to-text) or AutoModelForCausalLM

• FuTransformer: FP8 prefill and INT4 decode for Llama-style inference on Hopper
  paperas frameworks and tools2026-10-16

  Baseline compiled implementation for full Llama-3.1-8B decode

• Welcome Beam: a 501B open model you can own, customize, and scale
  articleas frameworks and tools2026-10-16

  Native Beam implementation (>=5.20.1) with pipeline, chat template, parse_response and DistributedConfig for expert parallelism

• Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta
  articleas frameworks and tools2026-10-15

  Hugging Face library providing MuseSparkForConditionalGeneration and MuseSparkProcessor; DistributedConfig sets tensor, expert and FSDP parallelism.

• Welcome FLUX.3 Image Klein in Diffusers
  articleas frameworks and tools2026-10-15

  Provides the Qwen3-VL text encoder

• Run fast, private agents on-device with LFM2.5-24B-A2B
  articleas classification.technologies2026-10-13
• EmbeddingGemma 2: an open, lightweight multimodal embedding model
  articleas frameworks and tools2026-10-06

  Used to serve the model efficiently

• zzzz7788990213-ops/EvoVLM
  repoas frameworks and tools2026-10-02

  Required for GPU inference in reproduction.

• [AINews] Meta Connect 2026: Muse glasses, voice, video, and Charm
  writeupas frameworks and tools2026-09-24It is written here as transformers (Hugging Face).

  Library supporting Nemotron 3 Diarization on day 0

• Transformers now runs llama.cpp quants
  articleas frameworks and tools2026-09-22

  Hugging Face library providing a foundation for model definition; now loads and runs GGUF quantized checkpoints, reusing ggml kernels via the kernels library and with an optimized generate loop for local inference.

• The current balance of power in open models
  articleas frameworks and tools2026-09-21

  Inference library accompanying open-weight model releases

• tokenizers v1: encode, decode and scaling, measured
  articleas frameworks and tools2026-09-21

  library that depends on tokenizers and is planned to receive its performance improvements after the release candidates stabilize

• Granite 4.2 LLMs: How They're Built
  articleas frameworks and tools2026-08-25

  Hugging Face library used to load the tokenizer and model and run inference (thinking mode, non-thinking mode, low-effort thinking, tool calling) for Granite 4.2.

• MakazhanAlpamys/Soup
  repoas frameworks and tools2026-08-14

  Backend library Soup trains against; a recipe can train a different result on this backend than on MLX

• Controlling Reasoning Effort in LLMs
  articleas frameworks and tools2026-07-18

  Library referenced for self-hosting GLM-5 using its open chat template to control turn-level thinking.

• Meta Muse Spark Open (816B-A42B) open-weights model card
  repoas frameworks and tools

  Hugging Face library; pipeline (image-text-to-text) and AutoModelForMultimodalLM/AutoProcessor to load and run the model

• google/embeddinggemma-2
  repoas frameworks and tools

  Library used to load google/embeddinggemma-2 via pipeline (feature-extraction) or AutoProcessor/AutoModel

• Qwen3.8-27B (Apache-2.0, multimodal dense, 262k ctx)
  repoas classification.technologies
• Tencent Hy4 preview (open-weight MoE + Sherry quantization)
  repoas frameworks and tools

  Library used to load and run tencent/Hy4-preview via pipeline or direct model/tokenizer loading

• DeepSeek-V4.1-Flash
  repoas frameworks and tools

  Library used to load and run DeepSeek-V4.1-Flash via pipeline or AutoModelForCausalLM

• MiMo-V2.6-Pro-RL
  repoas frameworks and tools

  Library to load and run MiMo-V2.6-Pro-RL via a pipeline or AutoModelForCausalLM

• MiMo-V2.6-Flash-RL (Xiaomi)
  repoas frameworks and tools

  Python library used to load and run MiMo-V2.6-Flash-RL via pipeline or AutoModelForCausalLM

In every workspace you can read 32