Qwen3.5-0.8B
The smallest published Qwen3.5 open-weights checkpoint with multimodal inputs, a hybrid decoder stack, and a 262,144-token native context window.
Qwen3.5-0.8B is the smallest Qwen3.5 checkpoint in the public release set. Official materials describe it as a post-trained open-weights model sized for prototyping, task-specific fine-tuning, and research or development work rather than as a large production default.
At a glance
- Family
- qwen
- Source type
- Open weights
- Released
- March 2026
- Authors
- Qwen Team
- Modalities
- Text, Image, Video
- Parameter count
- 0.8 billion parameters
- Context length
- 262,144 tokens
What It Is
Qwen3.5-0.8B is an open-weights model from the Qwen Team and the smallest published member of the Qwen3.5 release line. The official Hugging Face model card identifies this repository as the post-trained checkpoint, names the 0.8B operating point, and lists image-and-text serving examples alongside text-only use. In the public reader path, Qwen3.5 sits between the earlier Qwen3 generation and the later Qwen3.6 release line: the Qwen3.5 announcement describes this generation as a native multimodal agent upgrade over Qwen3, while Qwen3.6 release materials position the next generation as a stability-focused coding and agentic-workflow upgrade released after Qwen3.5.Inputs And Outputs
The official Hugging Face model card documents text and image prompts that produce text outputs, with serving examples for both modalities. The same checkpoint-specific source also includes a Video Input example and vision-language benchmark tables that evaluate video understanding, so video belongs here only because that card publishes it for Qwen3.5-0.8B. The card also documents a native 262,144-token context length as the default operating point.Architecture
The official model card's architecture overview names a hybrid language stack built from repeated Gated DeltaNet blocks paired with feed-forward layers, with periodic Gated Attention blocks also paired with feed-forward layers. Official serving notes also refer to a vision encoder that can be skipped for text-only deployment. The diagram below shows the high-level decoder flow—Gated DeltaNet as the dominant repeated path, Gated Attention at periodic depths, feed-forward blocks after each sequence step, and RoPE on the published attention path—without turning unpublished block counts into extra claims.Output
ProbabilitiesSoftmaxOutput
LinearTransformer architectureHybrid sequence sublayer container for Gated DeltaNet and Gated Attention pathsFeed-forward sublayer containerN×RMSNormGated
DeltaNetGated
AttentionRMSNormFeed
ForwardEmbeddingRoPEInput
EmbeddingInput
TokensSoftmax to Output
ProbabilitiesOutput
Linear to SoftmaxRMSNorm to Output
LinearFeed
Forward to RMSNormRMSNorm to Feed
ForwardRMSNorm to RMSNormGated
DeltaNet to RMSNormGated
Attention to RMSNormEmbedding to Gated
DeltaNetEmbedding to Gated
AttentionEmbedding to RMSNormRoPE to Gated
AttentionInput
Embedding to EmbeddingInput
Tokens to Input
Embedding
Important Modules
The registry-backed modules below are the smallest set worth opening first when you want to connect this page to the site's module reference. Rotary position encoding appears in the published Gated Attention head configuration on the official model card. Attention and context-window glossary pages explain the hybrid decoder and 262,144-token operating point; multimodal-model and modality glossary pages cover the published image and video inputs; the inference-engine system page is the serving path readers typically follow after prototyping here.Training
Official materials distinguish two public checkpoints at this size. Qwen/Qwen3.5-0.8B-Base is the pre-trained-only release intended for fine-tuning and in-context learning rather than direct chat use. Qwen/Qwen3.5-0.8B is the post-trained checkpoint that this page treats as the canonical reader route. Both checkpoints share the same published architecture overview on their Hugging Face model cards.Training regimes
No training regimes listed yet.
Linked papers
No linked paper pages listed yet.