# Hugging Face — Native-speed vLLM transformers modeling backend

- Company: Hugging Face (huggingface.co)
- Announced: 2026-07-08T00:00:00+00:00
- Category: capability-change
- Subject: Platform
- Models affected: Qwen3-4B, Qwen3-32B, Qwen3-235B-A22B-FP8
- Source: https://huggingface.co/blog/native-speed-vllm-transformers-backend
- Record: https://forck.live/items/1460-native-speed-vllm-transformers-modeling-backend

The transformers vLLM backend now achieves native or better throughput compared to custom vLLM implementations for many LLM architectures, using torch.fx and AST to apply runtime fusions.

## Evidence

Verbatim from https://huggingface.co/blog/native-speed-vllm-transformers-backend:

> The transformers vLLM backend is now as fast (or faster) than custom vLLM implementations for many LLM architectures.

---

Record: https://forck.live/items/1460-native-speed-vllm-transformers-modeling-backend
Catalogue: https://forck.live/llms.txt
Feed: https://forck.live/feed.md