LoginSign Up
← Back to all sparks
T

Transformers

AI-ASSISTANTS
Velocity5.0

Hugging Face library providing thousands of pretrained models for NLP, vision, and audio.

huggingface.co/docs/transformers ↗

Transformers keeps its model-a-release cadence, adding Kimi K2.5-2.7 and MiniMax/Diffusion variants

transformersopen-weight-modelsmodel-supportvllm-syncrelease-cadence
◆Current state
Transformers ships on a fast point-release train where nearly every minor version lands one or more new model architectures and the patch releases in between carry fixes — often to keep vLLM in sync. The v5.10-v5.13 window added Kimi K2.5/2.6/2.7, MiniMax-M3-VL, DiffusionGemma, Gemma4 Unified, and Cohere Command A+ (MoE), with several yank-and-republish hiccups along the way.
◆Where it's heading
The library continues as the reference implementation the open-weight ecosystem targets: model vendors upstream their architectures here on release day, and downstream serving stacks (vLLM) chase compatibility. The recurring patch releases syncing with vLLM and fixing conversion regressions show integration load is now as much of the work as new-model support itself.
◆Prediction
Expect the same rhythm to hold — a steady stream of minor releases each folding in the latest open-weight models, interleaved with vLLM-sync patch releases. No directional shift is visible in these entries.

◆Recent moves

  1. 2d ago

    v5.13.0 adds Kimi K2.5, 2.6, and 2.7 architectures

    Adds the Kimi K2.5-2.7 family — open-source multimodal agentic models aimed at long-horizon coding — continuing the pattern of upstreaming new open-weight architectures on release. A meaningful capability add for users, but in-line with the library's model-per-release norm rather than a directional shift.

    View source ↗
  2. 20d ago

    v5.12.1: PEFT lower-bound bump and Mistral tokenizer fix

    A patch release raising the PEFT lower bound and fixing auto-tokenizer resolution for the Mistral tokenizer. Housekeeping to keep the serving path working, no user-facing new capability.

    View source ↗
  3. 20d ago

    v5.10.3: vLLM-sync fixes and InternVL/processor patches

    A patch release of fixes needed to keep vLLM in sync with transformers, plus InternVL and ProcessorMixin token-id corrections. Integration maintenance, not a feature.

    View source ↗
  4. 23d ago

    v5.12.0 adds MiniMax-M3-VL vision-language model

    Adds MiniMax-M3-VL, the vision-language member of the MiniMax-M3 family pairing a CLIP-style tower with a mixed dense/sparse MoE backbone. Another new-architecture add consistent with the library's model-per-release train.

    View source ↗
  5. 25d ago

    v5.11.0 adds DiffusionGemma

    Adds DiffusionGemma, an encoder-decoder architecture built to cut the sequential decoding bottleneck of standard causal LMs. Extends model coverage in line with the ongoing cadence.

    View source ↗
  6. 1mo ago

    v5.10.2: fixes CLIP model conversion regression

    A patch fixing a conversion bug affecting CLIP-derived models such as SAM3. A correctness fix users should take, but no new capability.

    View source ↗