AWS AI Updates: September 15, 2026

1. Gemma 4 31B Landed on JumpStart in Both Full Precision and a 4-Bit NVIDIA Quantization

AWS. SageMaker JumpStart added Google DeepMind’s Gemma-4-31B-it-assistant alongside NVIDIA’s NVFP4-quantized variant of the same model. The assistant-tuned version takes text and image input, carries a 256K-token context window, covers more than 140 languages, and interleaves local sliding-window attention with full global attention. AWS says it ranks #3 among open models on the Arena AI text leaderboard. The NVFP4 build drops memory to roughly 18.5 GB, a 68% reduction, and runs about 2.5x faster while retaining 97-99% of base model quality. Source

2. Qwen3.6-35B-A3B Arrived in NVFP4 With a 262K Context Window

AWS. The NVIDIA-quantized variant of Alibaba’s Qwen3.6-35B-A3B is now deployable from JumpStart. It is a Mixture-of-Experts model with 35B total parameters activating 3B per token across 8 of 256 experts, supporting a 262K-token context extendable to roughly 1M through YaRN scaling. AWS highlights preserved thinking across conversation turns, multi-token prediction, and tool calling for multi-step agent pipelines. The same announcement added Wan2.1-T2V-1.3B-Diffusers, a 1.3B text-to-video model that produces a 5-second 480p clip on an RTX 4090 in about 4 minutes using 8.19 GB of VRAM. Source

3. Ministral 3 Brought Vision-Capable Edge Models Down to 8 GB of VRAM

AWS. Mistral AI’s Ministral-3-3B-Instruct-2512 and Ministral-3-8B-Instruct-2512 are now on JumpStart. The 3B variant pairs a 3.4B language model with a 0.4B vision encoder and fits in 8 GB of VRAM at FP8 while supporting a 256K-token context window and multilingual instruction following across dozens of languages. Both are positioned for edge and resource-constrained deployment rather than datacenter serving. Source

4. JumpStart Added Speech, Bilingual Agentic, and Structure Prediction Models

AWS. Three more models went live: IBM’s granite-speech-4.1-2b, Kakao’s kanana-2-30b-a3b-instruct, and the OpenFold Consortium’s OpenFold3. The granite speech model handles ASR and bidirectional speech translation across six languages at a 5.33% word error rate with a real-time factor near 231, released under Apache 2.0. Kanana-2 is an MoE model activating 3B of 30B parameters per forward pass, using Multi-head Latent Attention and supporting 128K tokens via YaRN. OpenFold3 covers all-atom structure prediction for proteins, DNA, RNA, and small-molecule ligands. Source

5. Glue Zero-ETL Now Tracks Which Integration Owns a Target Table

AWS. Glue zero-ETL integrations detect table property conflicts and record integration ownership across Amazon S3 Tables and SageMaker Lakehouse catalogs. Two integrations can no longer be pointed at the same target table silently. Attempting to create or modify an integration whose table properties belong to another integration now surfaces the owning integration and prompts for a different target. The feature is available in all Commercial and GovCloud (US) regions where Glue zero-ETL runs. Source