Amazon SageMaker JumpStart Expands Foundation Model Portfolio with Frontier Releases from Google DeepMind, NVIDIA, Mistral AI, and Alibaba

⚡ EXECUTIVE SUMMARY (30-Second Brief)

• The Core Event: Amazon Web Services (AWS) has expanded its Amazon SageMaker JumpStart catalog with six advanced open foundation models spanning Mixture-of-Experts (MoE) reasoning, 4-bit FP4 quantized acceleration, compact edge vision, and lightweight text-to-video generation [Source 1, Source 2, Source 3].
• Key Technical Metrics:
  • Qwen3.6-35B-A3B-NVFP4: Mixture-of-Experts architecture with 35B total parameters, 3B active per token (8 of 256 experts), and a 262K-token context window extendable to ~1M tokens via YaRN scaling [Source 1].
  • Gemma-4-31B-IT-NVFP4: FP4 quantization via NVIDIA ModelOpt cuts memory footprint to ~18.5 GB (68% reduction) and delivers ~2.5x faster inference while maintaining 97–99% base model quality [Source 3].
  • Wan2.1-T2V-1.3B-Diffusers: 1.3B parameter diffusion transformer requiring 8.19 GB VRAM, capable of rendering a 5-second 480p video on an NVIDIA RTX 4090 in ~4 minutes [Source 1].
  • Ministral 3 Family: Compact vision-language models fitting in 8GB (3B variant) and 12GB (8B variant) VRAM in FP8 precision [Source 2].
• Immediate Ecosystem Impact: AWS enterprise customers can now deploy high-efficiency, specialized open-weights models—ranging from ultra-lightweight edge endpoints to long-context agentic reasoning pipelines—directly via the SageMaker Console or Python SDK [Source 1, Source 2, Source 3].

---

🔬 CORE REPORT & STRATEGIC NARRATIVE

1. NVIDIA-Quantized Frontier Models: FP4 Efficiency & Active MoE Routing

To address the escalating compute costs of serving large-scale models, NVIDIA has optimized and quantized top-tier architectures using its ModelOpt framework:

• Qwen3.6-35B-A3B-NVFP4: Derived from Alibaba's Qwen3.6-35B-A3B, this Mixture-of-Experts (MoE) model features 35 billion total parameters, but dynamically routes tokens to activate only 3 billion parameters (8 out of 256 experts) per token. Quantized to NVFP4, it native-supports a 262K-token context window that can be scaled to ~1M tokens using YaRN context extension. It preserves chain-of-thought thinking across multi-turn conversations, multi-token prediction, and multi-step tool calling for agentic workflows [Source 1].
• Gemma-4-31B-IT-NVFP4: Based on Google DeepMind’s flagship dense 31B assistant architecture, this model utilizes NVIDIA ModelOpt 4-bit FP4 quantization to reduce the memory footprint to ~18.5 GB—a 68% reduction compared to the base model. The quantized release achieves approximately 2.5x faster inference speeds while retaining 97% to 99% of the full-precision model's benchmarks [Source 3].

2. Compact Multimodal Edge Models: Mistral’s Ministral 3 Family

For resource-constrained endpoints and low-latency workloads, Mistral AI’s Ministral 3 lineup brings native multimodal capabilities to sub-10B parameter deployments:

• Ministral-3-3B-Instruct-2512: Composed of a 3.4B language model paired with a 0.4B vision encoder, this model fits inside 8GB of VRAM when deployed in FP8. Released under the open Apache 2.0 license, it supports a 256K-token context window, multilingual instruction following across dozens of languages (including English, French, Spanish, German, Chinese, Japanese, Korean, and Arabic), system-prompt adherence, and structured JSON tool-calling [Source 2].
• Ministral-3-8B-Instruct-2512: Combining an 8.4B language model with a 0.4B vision encoder, this model runs within 12GB of VRAM in FP8. It introduces an interleaved sliding-window attention pattern to lower memory pressure during long-sequence generation, matching the operational scope of larger models while offering stronger reasoning than the 3B variant [Source 2].

3. Open Multimodal & Video Generation Powerhouses

Rounding out the deployment catalog are Google DeepMind’s flagship dense model and Alibaba’s lightweight text-to-video diffusion transformer:

• Gemma-4-31B-it-assistant: Designed for high-tier multimodal reasoning, coding, and autonomous agents, this 31B dense model ranks #3 among open models on the Arena AI text leaderboard. It processes text and image inputs (treating video as frame sequences), supports a 256K-token context window, handles over 140 languages, and uses a hybrid attention architecture that interleaves local sliding-window attention with full global attention [Source 3].
• Wan2.1-T2V-1.3B-Diffusers: Built on a diffusion transformer architecture paired with a Video Variational Autoencoder (VAE), this 1.3B-parameter model enables local text-to-video generation on consumer hardware. Operating within 8.19 GB of VRAM, it generates physics-consistent, 5-second 480p video clips on an NVIDIA RTX 4090 GPU in approximately 4 minutes [Source 1].

---

📊 VISUAL INFOGRAPHIC / FLOWCHART

VISUAL FLOWCHART / INFOGRAPHICDIAGRAM