Overview

Purpose and Scope

This page explains how the Transformers documentation overview is expressed through individual model documentation pages, rather than through a single central implementation file. The requested sources are all docs/source/en/model_doc pages, so the useful reader problem is understanding the pattern they share: each page introduces a model family, names the intended tasks and modalities, shows the loading path, then delegates detailed API signatures to generated reference blocks. Read this as a map for navigating model documentation and for recognizing where high-level library concepts such as pipelines, Auto classes, processors, generation, attention backends, and quantization appear in concrete model pages.

Sources: docs/source/en/model_doc/qwen3_next.md, docs/source/en/model_doc/llava_next_video.md, docs/source/en/model_doc/llava_next.md, docs/source/en/model_doc/qwen3_5_moe.md, docs/source/en/model_doc/qwen3_5.md, docs/source/en/model_doc/qwen3_asr.md

The selected pages also show how the official documentation surface stays consistent while covering very different model families. Qwen3-Next is presented as a long-context, sparse foundation model with hybrid attention, mixture-of-experts behavior, and causal language modeling examples. LLaVA-NeXT and LLaVA-NeXT-Video present multimodal image and video understanding through pipeline and AutoModel flows. Qwen3.5 and Qwen3.5 MoE extend the Qwen family with dense and sparse variants, while Qwen3 ASR covers speech recognition and forced alignment. Together, they demonstrate that the docs overview is model-family oriented but anchored in common Transformers primitives.

Relevant Source Files

docs/source/en/model_doc/qwen3_next.md - Documents the Qwen3-Next family, including its Overview section, architectural summary, chat-style generation example, and autodoc entries for config, base model, causal LM, and classification or QA heads.
docs/source/en/model_doc/llava_next_video.md - Documents LLaVA-NeXT-Video, including the model title, attention badges, overview narrative, relationship to LLaVA-NeXT, AnyRes video framing, and video-understanding context.
docs/source/en/model_doc/llava_next.md - Documents LLaVA-NeXT, including image-text usage through both pipeline and AutoProcessor plus LlavaNextForConditionalGeneration, and notes on quantization for memory reduction.
docs/source/en/model_doc/qwen3_5_moe.md - Documents Qwen3.5 MoE, including sparse expert architecture, checkpoint examples, pipeline and AutoModel quickstarts, router-logit training notes, and long-context configuration guidance.
docs/source/en/model_doc/qwen3_5.md - Documents dense Qwen3.5 and Qwen3.6 variants, including native multimodal training, hybrid attention layout, quickstart examples, optional kernel dependencies, and model-family relationships.
docs/source/en/model_doc/qwen3_asr.md - Documents Qwen3 ASR, including speech-to-text architecture, available ASR and forced-aligner checkpoints, supported languages, and processor-based transcription usage.

Documentation Structure

A model documentation page usually starts by establishing provenance and scope before it teaches usage. The Qwen3-Next page states its contribution date, then opens with an ## Overview that summarizes architectural innovations such as hybrid attention, high-sparsity MoE, multi-token prediction, and training or inference efficiency. The Qwen3 ASR page similarly names the model, explains that it combines a Whisper-style audio encoder with a Qwen3 language-model decoder, and distinguishes transcription from forced alignment. This front-loaded context helps readers decide whether a model family matches their task before they copy any code.

Sources: docs/source/en/model_doc/qwen3_next.md, docs/source/en/model_doc/qwen3_asr.md

The LLaVA pages show the same structure applied to vision-language and video-language models. llava_next.md begins with the model name, capability statement, checkpoint collection guidance, and usage options. Its examples demonstrate both the high-level pipeline abstraction for image-text-to-text and a lower-level AutoProcessor plus model class path for users who need direct control over prompts, tensors, and generation. llava_next_video.md builds on that earlier page by explaining how AnyRes image handling transfers to videos as sets of frames and why supervised fine-tuning improves video understanding.

Sources: docs/source/en/model_doc/llava_next.md, docs/source/en/model_doc/llava_next_video.md

Qwen3.5 and Qwen3.5 MoE illustrate how related variants are documented without forcing readers to infer compatibility. The dense page explicitly frames Qwen3.5 as a natively multimodal family trained on interleaved text, image, and video tokens, and it points MoE readers to the separate sparse page. The MoE page then defines the sparse-expert variant, names active and total-parameter checkpoints, and warns that shapes differ from dense checkpoints. That split is important: the overview layer is conceptual, but it still protects users from incompatible weights, missing router logits, and incorrect assumptions about active compute.

Sources: docs/source/en/model_doc/qwen3_5.md, docs/source/en/model_doc/qwen3_5_moe.md

System-to-Code Mapping

The repository path pattern docs/source/en/model_doc/*.md is the documentation-side counterpart of model families exposed through Transformers APIs. In these pages, concrete examples use public loading primitives such as AutoTokenizer.from_pretrained, AutoProcessor.from_pretrained, AutoModelForCausalLM.from_pretrained, AutoModelForMultimodalLM.from_pretrained, LlavaNextForConditionalGeneration.from_pretrained, and task-specific pipeline calls. The pages are therefore not implementation modules themselves; they are the published contract that tells users which public entry point to choose, which checkpoint identifier to pass, and which preprocessing object should prepare model inputs.

Sources: docs/source/en/model_doc/qwen3_next.md, docs/source/en/model_doc/llava_next.md, docs/source/en/model_doc/qwen3_5.md, docs/source/en/model_doc/qwen3_5_moe.md, docs/source/en/model_doc/qwen3_asr.md

The [[autodoc]] blocks in the Qwen3-Next page show another important mapping. Instead of manually repeating every method signature, the page lists documented classes such as Qwen3NextConfig, Qwen3NextModel, Qwen3NextForCausalLM, Qwen3NextForSequenceClassification, Qwen3NextForQuestionAnswering, and Qwen3NextForTokenClassification, with forward methods attached where relevant. The doc builder expands those blocks from the Python API. This keeps the model page readable while preserving a route to precise generated reference documentation, matching the broader docs practice of combining narrative examples with generated class documentation.

Sources: docs/source/en/model_doc/qwen3_next.md

Execution Flow for Readers

A practical reading flow starts with modality. If the input is text-only generation and the user wants a long-context sparse model, Qwen3-Next or Qwen3.5 MoE are the relevant pages. If the input is image plus text, LLaVA-NeXT demonstrates the multimodal message format, including content items with type values such as image and text. If the input is audio, Qwen3 ASR shows a processor method, apply_transcription_request, that prepares a transcription request before generate is called. The overview documentation therefore teaches workflow selection before optimization details or class internals.

Sources: docs/source/en/model_doc/qwen3_next.md, docs/source/en/model_doc/llava_next.md, docs/source/en/model_doc/qwen3_5_moe.md, docs/source/en/model_doc/qwen3_asr.md

After choosing the model family, readers should compare the high-level and low-level examples. Pipeline examples are shortest and show the supported task label, such as text-generation or image-text-to-text, plus checkpoint and device_map selection. AutoModel examples expose the processor or tokenizer, chat template application, tensor conversion, model.generate, and decoding. This distinction is central to the docs overview: pipelines are convenient for fast inference, while direct model calls are better when the task requires custom prompt formatting, multimodal tensor handling, output slicing, or training-adjacent experimentation.

Sources: docs/source/en/model_doc/llava_next.md, docs/source/en/model_doc/qwen3_5.md, docs/source/en/model_doc/qwen3_5_moe.md, docs/source/en/model_doc/qwen3_asr.md

API Components and Options

The shared public contract across these pages is from_pretrained(checkpoint, ...) plus a preprocessing object and a model execution call. Text pages use tokenizers and causal language models; multimodal pages use processors because they must combine text with image, video, or audio inputs; generation pages call generate and then decode newly generated ids. Model-specific options appear where they matter: Qwen3.5 pages discuss optional kernels, use_kernels=True, hybrid layer_types, YaRN rope scaling, and context length behavior; the MoE page discusses output_router_logits, router_aux_loss_coef, and router logits in outputs.

Sources: docs/source/en/model_doc/qwen3_5.md, docs/source/en/model_doc/qwen3_5_moe.md, docs/source/en/model_doc/qwen3_asr.md

Area	Concrete docs signal	Reader action
Text generation	`AutoTokenizer`, `AutoModelForCausalLM`, `generate`	Load a checkpoint, tokenize prompts, generate ids, decode text.
Image-text	`pipeline(task="image-text-to-text")`, `AutoProcessor`, `LlavaNextForConditionalGeneration`	Choose convenience pipeline or explicit processor/model flow.
Video understanding	LLaVA-NeXT-Video overview and AnyRes framing	Treat videos as frame sequences in the multimodal model context.
Sparse MoE	`output_router_logits`, `router_aux_loss_coef`, router logits	Enable router outputs during training or fine-tuning to avoid expert collapse.
Speech recognition	`AutoProcessor`, `AutoModelForMultimodalLM`, `apply_transcription_request`	Prepare audio requests, run generation, decode ASR output.

Implementation and Publishing Details

These Markdown files use documentation-builder syntax in addition to ordinary Markdown. The source comments warn that the files contain syntax similar to MDX, and the pages use custom elements such as hfoptions, hfoption, badges, tips, and [[autodoc]] blocks. That matters for contributors because a page may not render fully in a plain Markdown viewer, even though it is the source of the published documentation. The correct mental model is that prose, examples, and metadata live beside doc-builder directives that connect the written guide to generated API references and interactive tabs.

The official documentation snippets reinforce two constraints that are visible in these model pages. First, optimization language should distinguish memory, speed, kernels, attention backends, quantization, caching, and parallelism instead of treating them as a single feature. The LLaVA-NeXT page links quantization to lower-precision weights and memory reduction, while Qwen3.5 discusses optional kernel paths and fallback behavior. Second, generated docstrings and model structure rules exist so model classes, configs, processors, and methods can be documented consistently. The source pages rely on that infrastructure by keeping human-authored overview text close to generated class references.

Sources: docs/source/en/model_doc/llava_next.md, docs/source/en/model_doc/qwen3_5.md, docs/source/en/model_doc/qwen3_next.md

Next Steps

Use this overview as an orientation layer, then move to the page that matches your immediate task. For quick inference, start with the pipeline examples in LLaVA-NeXT, Qwen3.5, or Qwen3.5 MoE and only drop to AutoModel code when you need control over prompts, tensors, or outputs. For multimodal work, compare AutoProcessor examples across image, video, and audio pages. For contribution work, preserve this structure: introduce the model, state variants and checkpoints, provide runnable examples, document model-specific caveats, and connect public classes through autodoc rather than duplicating generated API reference text.

Related pages: quickstart, pipelines-guide, auto-classes-and-model-loading, processors, text-generation, multimodal-tasks, audio-tasks, quantization-overview, kernels, contributing

Ask a question...