JA EN
shiichan

From Long-Context Coding Agents to Video Generation — Qwen3.6 and Wan2.1 Land on SageMaker JumpStart!

AWS What's New ★★★★ View Markdown
#aws#ai

Hi everyone, I'm Shiichan! Today's news covers two very different new models on SageMaker JumpStart: one for coding agents, one for video generation.

AWS What's New aws.amazon.com

What was announced?

AWS What's New announced that two models are now available on Amazon SageMaker JumpStart: Alibaba's Qwen3.6-35B-A3B-NVFP4 (an NVIDIA-quantized variant) and Wan2.1-T2V-1.3B-Diffusers. One targets agentic coding, the other text-to-video generation — two completely different domains.

Why it matters

Agentic coding that keeps calling tools while holding onto a long context, and video generation that actually runs on consumer GPUs, are both the kind of technology you want to try on your own hardware right away. Now both are a few clicks away on JumpStart.

What changes

Here's what each model brings:

  • Qwen3.6-35B-A3B-NVFP4: Optimized for agentic coding, multimodal reasoning, and long-context understanding. It's a Mixture-of-Experts model with 35 billion total parameters, activating only 8 of 256 experts (3 billion parameters) per token. Its context window is 262K tokens, extendable to roughly 1 million with YaRN scaling. Quantized to NVFP4 with NVIDIA's ModelOpt framework, it preserves reasoning across conversation turns, supports multi-token prediction, and handles tool calling, all with a reduced memory footprint.
  • Wan2.1-T2V-1.3B-Diffusers: A text-to-video generation model that runs on consumer-grade hardware. It combines a diffusion transformer with a novel video VAE (Variational Autoencoder) to produce physics-consistent video. At 1.3 billion parameters, it needs just 8.19 GB of VRAM, and can generate a 5-second 480p clip on an RTX 4090 in about 4 minutes.

Dive Deep

Qwen3.6-35B-A3B-NVFP4 is well suited to multi-step agent workflows that need to carry reasoning from earlier turns forward while repeatedly calling tools. Wan2.1-T2V-1.3B-Diffusers, meanwhile, stands out for its relatively small 1.3B parameter size, which brings video generation within reach of a high-end but still consumer GPU like the RTX 4090 — lowering the infrastructure bar if you just want to experiment with video generation. Deploy either from the SageMaker JumpStart model catalog in a few clicks, or via the Python SDK.

Wrap-up

  • Qwen3.6-35B-A3B-NVFP4 (agentic coding) and Wan2.1-T2V-1.3B-Diffusers (text-to-video) are now on SageMaker JumpStart
  • Qwen3.6 supports a 262K-token context (extendable to ~1M) for long-context agent workflows
  • Wan2.1 runs in 8.19GB of VRAM and generates video even on an RTX 4090
  • Worth checking out whether you're building long-context agents or want to try video generation on a consumer GPU

Related posts