Verified · Jul 8, 2026

HF transformers 7/3 v5.13.0: 8 new models + HfExporters, ASR / vision / long context all in

4 sources

huggingface/transformers v5.13.0 (July 3) adds 8 new models in one release: LLM side Kimi K2.5/2.6/2.7 (Moonshot's multimodal agent for long-horizon coding, front-end design, Rust/Go/Python swarm orchestration), Xiaomi MiMo-V2-Flash (27T-token MoE, native 32k context, extendable 256K), Zyphra ZAYA1 (760M active / 8.4B total MoE, Compressed Convolutional Attention), OpenBMB MiniCPM3-4B (MLA from DeepSeek-V2 + SwiGLU); vision/video: Google DeepMind VideoPrism (36M video-caption pairs + 582M noisy clips pretraining) + NVIDIA RADIO (CLIP/DINOv2/SAM distilled into a single variable-resolution ViT); ASR: NVIDIA Nemotron 3.5 ASR + Nemotron ASR Streaming (80/160/560/1120 ms chunk configurable) + Qwen3 ASR (Whisper-style encoder + Qwen3 decoder + forced-aligner); plus the HfExporters subsystem (DynamoExporter / OnnxExporter / ExecutorchExporter subclasses + prefill/decode auto-split + layer / mask / cache standardization).

Why now

v5.13.0 is the first-week-of-July max-capacity expansion of the model library (8 new models, LLM/ASR/vision three directions in one release), with the long-promised HfExporters subsystem landing alongside. Creators can do 'one release that covers the first week of July research-side dynamics.'

Why it is worth publishing

Each of the 8 new models corresponds to a different direction (Kimi agent / MoE long context / compressed attention / MLA dense small / video encoder / multi-teacher ViT / multilingual ASR / streaming ASR / Whisper+Qwen3 ASR) — perfect for a 'one release, one read' horizontal comparison.

Evidence basis

v5.13.0 is a major transformers release, stable mid-to-upper heat. Among the 8 new models, Kimi / VideoPrism / RADIO are July's relatively high-attention directions; combined release heat is stable.

“'transformers 7/3 dropped 8 new models and unified ONNX/ExecuTorch export in the same release — one release, one read of July's first week on the research side.'”

Angle

Frame v5.13.0 as 'one release, one read of July research-side dynamics' — creators make 'LLM side (Kimi + MiMo + ZAYA1 + MiniCPM3) / vision side (VideoPrism + RADIO) / ASR side (Nemotron 3.5 / Streaming / Qwen3) all in, plus HfExporters as the engineering-side export target' multi-column comparison.

Format

Long-form explainer

Demo idea

10-minute v5.13.0 demo: run Kimi K2.5 multimodal agent for a file read/write, run Qwen3 ASR on a phone recording, run NVIDIA RADIO on an image, then switch to HfExporters OnnxExporter to export Kimi. Shows 'models + standardized export' full flow.

Platform notes

Per-model benchmark numbers (WER, accuracy, latency, throughput), training dataset references, per-PR migration notes not in the captured release notes (medium risk). HfExporters specific supported model coverage, export-vs-eager performance numbers, downstream-library migration steps not given (medium risk). Don't fabricate numbers from memory.

Usable claims

huggingface/transformers v5.13.0 (released 2026-07-03) adds 8 new models: Kimi K2.5/2.6/2.7 (multimodal agentic architecture from Moonshot for long-horizon coding, front-end design, and swarm-based task orchestration across Rust, Go, and Python); Xiaomi MiMo-V2-Flash (27T-token MoE with native 32k context and extendable 256K window); NVIDIA Nemotron 3.5 ASR (multilingual) and Nemotron ASR Streaming (English, cache-aware FastConformer-RNNT with configurable 80/160/560/1120 ms chunk sizes); Qwen3 ASR (Whisper-style encoder + Qwen3 decoder with forced-aligner head); Zyphra ZAYA1 (760M active / 8.4B total MoE, Compressed Convolutional Attention); Google DeepMind VideoPrism (general-purpose video encoder pretrained on 36M video-caption pairs and 582M noisy clips); NVIDIA RADIO (distills CLIP, DINOv2, and SAM into a single variable-resolution ViT); OpenBMB MiniCPM3-4B (MLA from DeepSeek-V2 + SwiGLU).
huggingface/transformers v5.13.0 (released 2026-07-03) introduces the HfExporters subsystem: a unified base class with DynamoExporter, OnnxExporter, and ExecutorchExporter subclasses giving a single API for PyTorch/ONNX/ExecuTorch export with automatic prefill/decode splitting for generative models; the release standardizes layer declarations, mask/cache construction, and hybrid-attention handling to make models cleanly ONNX-, torch.export-, and ExecuTorch-exportable.

Evidence pipeline

From the news

Breakdown

8 new models + HfExporters read individually becomes a number-reading script. This explainer walks through the 'one release, one read of July research-side dynamics' frame — group the 8 models into LLM / vision / ASR, then add HfExporters as a separate 'engineering-side export target' section, so creators make 'one release, one read of the week on the research side' content.

Sources

Risks

Pin links to each source; quote only what the captured summary states; do not paraphrase specific GitHub PR numbers, HF blog body details, features-article specifics, performance metrics, or migration details beyond what is stated.

Demo ideas

Run Kimi K2.5 swarm orchestration on Rust/Go/Python files to show multi-language agent collaboration.
Compare Qwen3 ASR and Nemotron ASR Streaming on the same phone recording, showing the forced-aligner timestamp head difference.
Use HfExporters to export MiniCPM3-4B to ONNX, run a generation, compare eager-mode vs exported-mode performance.