What's New

AWS has added four new Qwen models to Amazon SageMaker JumpStart, announced on May 4, 2026: Qwen3.5-27B-FP8, Qwen3.6-35B-A3B, Qwen3.5-0.8B, and Qwen3.5-2B. These models span a wide capability spectrum, from a large multimodal vision-language model supporting up to 1 million token context windows, to a compact sub-1B parameter model suited for edge deployments. Together, they address use cases including agentic coding, multimodal reasoning, multilingual communication across 200+ languages, and on-device inference.

How It Works

• All four models are hosted and served through Amazon SageMaker JumpStart, AWS's model hub that provides one-click deployment of pre-trained foundation models onto managed SageMaker inference endpoints.
• Users can deploy any of the four models either through the SageMaker Studio UI with a few clicks or programmatically via the SageMaker Python SDK.
• Qwen3.5-27B-FP8 uses FP8 quantization to reduce memory footprint while preserving the reasoning quality of a 27B-parameter multimodal model capable of processing images, video, and text.
• Qwen3.6-35B-A3B employs a Mixture-of-Experts (MoE) architecture with only 3 billion active parameters at inference time despite having 35 billion total parameters, making it computationally efficient for agentic coding workflows.
• The two smaller models (0.8B and 2B) are standard dense multimodal architectures optimized for low-latency, low-resource deployments including fine-tuning scenarios and edge inference.

Why It's Important

• This release significantly broadens the range of Qwen models available natively on AWS infrastructure, enabling teams to select a model precisely matched to their compute budget and task requirements without managing custom model imports or container configurations.
• The inclusion of a 1-million-token context model (Qwen3.5-27B-FP8) is particularly impactful for enterprise use cases involving long documents, large codebases, or extended multi-turn agent interactions.
• The MoE-based Qwen3.6-35B-A3B offers a compelling cost-performance trade-off for coding copilot and agentic applications, since only 3B parameters are active per forward pass, dramatically reducing inference cost relative to its total parameter count.
• The sub-2B models lower the barrier for rapid prototyping, fine-tuning experiments, and edge deployments where GPU memory is constrained.

How It's Different

• Prior to this announcement, SageMaker JumpStart's Qwen model catalog was more limited in scope, lacking this particular combination of MoE-based coding specialists, FP8-quantized large multimodal models, and ultra-compact edge-ready variants.
• Compared to deploying these models manually—by pulling weights from Hugging Face, building custom Docker containers, and configuring SageMaker endpoints—JumpStart eliminates significant undifferentiated heavy lifting.
• Relative to Amazon Bedrock's managed model offerings, JumpStart provides more flexibility for custom fine-tuning, direct endpoint configuration, and access to open-weight models that may not be available through Bedrock's fully managed API.
• The FP8 quantization on the 27B model is also noteworthy compared to standard BF16 or INT8 deployments, offering a better balance of throughput and accuracy for production workloads.

When to Prefer It

• Choose Qwen3.5-27B-FP8 when your application requires multimodal understanding (images, video, text), very long context windows (up to 1M tokens), or broad multilingual support across 200+ languages, and you need a production-grade model with reduced memory overhead via FP8 quantization.
• Prefer Qwen3.6-35B-A3B for agentic coding pipelines, repository-level code reasoning, or frontend development copilots where inference cost efficiency is critical—the MoE architecture delivers near-large-model quality at a fraction of the active compute cost.
• Use Qwen3.5-0.8B for edge deployments, on-device inference, or rapid fine-tuning experiments where GPU memory is severely constrained.
• Qwen3.5-2B is the right choice for moderate-compute environments that need a balance between capability and resource consumption, such as conversational AI prototypes or lightweight visual understanding tasks.
• All four are preferable over self-managed deployments when operational simplicity, fast time-to-deployment, and AWS-native integration (IAM, VPC, CloudWatch) are priorities.

Availability

• All four Qwen models are generally available (GA) in Amazon SageMaker JumpStart as of May 4, 2026.
• Deployment is supported through both Amazon SageMaker Studio and the SageMaker Python SDK.
• The announcement does not specify regional restrictions, but SageMaker JumpStart availability is generally tied to regions where SageMaker is fully supported; users should verify availability in their specific AWS region via the SageMaker console or documentation.
• No explicit limitations on instance types are stated in the announcement, though practical constraints apply: FP8 and large MoE models will require GPU instances with sufficient VRAM (e.g., ml.g5 or ml.p4/p5 family), while the 0.8B and 2B models can likely run on smaller GPU or potentially CPU-backed instances.
• Standard SageMaker JumpStart pricing applies based on the underlying compute instance selected for the endpoint.