LLMs

Amazon SageMaker JumpStart now hosts NVIDIA-optimized Qwen3.6-35B-A3B-NVFP4 and Alibaba's Wan2.1-T2V-1.3B-Diffusers models. The Qwen variant is a Mixture-of-Experts architecture with 35B total parameters but only 3B active per token, featuring NVFP4 quantization for agentic coding and long-context tasks. Wan2.1 provides lightweight text-to-video generation capabilities directly to AWS customers.

• Qwen3.6 uses MoE with 3B active parameters per token, reducing inference compute vs full 35B dense models.
• NVFP4 quantization on Qwen3.6 optimizes memory and speed for agentic coding and multimodal reasoning.
• Wan2.1 offers a 1.3B parameter diffusion model for efficient text-to-video generation workflows.
• Both models are immediately deployable via SageMaker JumpStart for scalable AI solution building.

Mistral AI's Ministral-3-3B and 8B instruct models are now available on Amazon SageMaker JumpStart. The 3B variant is optimized for ultra-lightweight edge deployment, combining a 3.4B language model with a 0.4B vision encoder. It requires only 8GB of VRAM in FP8 format while supporting a 256K-token context window.

• Ministral-3-3B fits in 8GB VRAM (FP8) for resource-constrained edge environments.
• Supports 256K-token context windows despite its compact model size.
• Integrates a 0.4B vision encoder for multimodal understanding capabilities.
• Expanded SageMaker JumpStart portfolio adds purpose-built foundation models.

Amazon SageMaker JumpStart now hosts Google DeepMind's Gemma-4-31B-it-assistant and NVIDIA's Gemma-4-31B-IT-NVFP4 models. These variants of the 31B dense architecture support multimodal reasoning, coding, and agentic workflows with a 256K context window. The availability includes both a standard assistant-tuned version and an NVIDIA-optimized quantized option for scalable deployment.

• SageMaker JumpStart adds Gemma-4-31B in full-precision and NVIDIA NVFP4 quantized forms.
• Models support multimodal inputs including text, images, and video frame sequences.
• Optimized for agentic workflows and coding tasks with a 256K context window.
• Provides enterprise-ready deployment paths for high-performance AI on AWS.
• NVIDIA variant offers quantization benefits for efficient inference scaling.

This paper introduces inference networks, a graph-based framework where nodes represent distinct LLMs and edges define conditional activation paths. The approach addresses the lack of principled strategies for combining expert models in ensemble or series configurations. By dynamically routing queries based on complexity, the system aims to balance inference costs with performance. It provides a structured method to leverage multiple models rather than relying on a single large model for all tasks.

• Uses graph structures to map conditional activation paths between different LLMs.
• Enables dynamic routing of simple queries to cheaper models and complex ones to powerful ones.
• Addresses the gap in systematic strategies for combining expert models in synergy.
• Offers a framework to optimize the trade-off between inference cost and model performance.

ZGCM-1 is a fully open 7B dense foundation model trained from scratch to handle math and agentic search tasks efficiently. It addresses parametric capacity limits by coupling deliberate internal thinking with active external tool use. The training recipe features an end-to-end high-efficiency design, including interleaved gated sliding-window and full attention mechanisms, and stable FP8 Muon optimization.

• Compact models can exceed parametric limits via internal reasoning + external tool use.
• Architecture uses interleaved gated sliding-window and full attention for 256K context.
• Training employs stable FP8 Muon optimizer for system and algorithmic efficiency.
• Progressive curriculum scales context from 16K to 256K during mid-training.
• Interaction traces are reformulated into Markov Decision Processes for better learning.

TimeThink addresses the inability of current timeseries multimodal LLMs to handle out-of-distribution compositional questions. The framework uses synthetic data to elicit explicit reasoning capabilities, moving beyond the implicit answers that fail in high-stakes domains. It specifically targets the gap where reinforcement learning models struggle with dynamic temporal patterns and complex queries.

• Current TS-MLLMs lack explicit reasoning for high-stakes applications like healthcare.
• TimeThink provides a synthetic framework to elicit compositional timeseries reasoning.
• Addresses failures in out-of-distribution queries that plague RL-based models.
• Enables better handling of dynamic temporal patterns through structured primitives.