DeepSeek's DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These three open-weight Mixture-of-Experts (MoE) language models, released under the MIT license, deliver frontier-class performance across reasoning, coding, and agentic workloads while activating only a fraction of their total parameters per token, enabling high efficiency at scale. These models address different AI challenges with specialized capabilities: DeepSeek-V4-Pro is designed for complex reasoning and high-quality text generation at flagship scale. This 1.6T-parameter MoE model activates 49B parameters per token and supports a 1M-token context window, making it well-suited for long-document analysis, multi-step reasoning, and demanding generation tasks that require maximum model capacity. DeepSeek-V4-Flash delivers strong performance at significantly lower compute cost as an efficient 284B-parameter MoE model activating 13B parameters per token. It supports the same 1M-token context window as V4-Pro, making it ideal for latency-sensitive and cost-conscious production deployments across a wide range of language tasks. DeepSeek-V4.1-Flash provides next-generation agentic efficiency with native multimodal support. Built on a novel Causal Encoder-Decoder (CED) architecture with a 552B-parameter backbone, it activates just 8B parameters during prefill and 16B during decode, supporting a 1M-token context window and 384K max output tokens. Its Compressed Sparse Attention 2 (CSA2) and FP4 KV caching reduce the global KV cache footprint to roughly 1/4 of V4-Flash, delivering substantially improved cost efficiency for agentic workloads, coding, and vision tasks. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

DeepReinforce AI's Ornith-1.0-9B and Ornith-1.0-35B models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models are part of the Ornith 1.0 family of open-source, self-scaffolding agentic coding models released under the MIT license, achieving state-of-the-art performance among open-source models of comparable size on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo, and OpenClaw. These models address different agentic coding challenges with specialized capabilities: Ornith-1.0-9B is a lightweight 9B-parameter dense model post-trained on Qwen 3.5 9B, designed for efficient single-GPU deployment. With a 256K-token context window and 32 decoder layers, it is well-suited for individual developers and small teams running agentic coding workflows on a single GPU — including laptops with consumer-grade hardware. Its self-improving training framework uses reinforcement learning to jointly optimize scaffold generation and solution rollouts, enabling the model to discover better search trajectories and generate higher-quality code solutions. Ornith-1.0-35B is a 35B-parameter Mixture-of-Experts (MoE) model post-trained on Qwen 3.5, delivering stronger agentic coding performance while remaining deployable on a single 24GB consumer GPU. Like its smaller sibling, it employs reinforcement learning to jointly optimize the scaffolds that drive coding tasks and the resulting solution rollouts, discovering superior code trajectories. Its MoE architecture activates only a subset of parameters per token, providing a strong balance of quality and efficiency for production agentic coding workloads. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

DeepSeek's DeepSeek-V4-Pro, DeepSeek-V4-Flash, and DeepSeek-V4.1-Flash models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These three open-weight Mixture-of-Experts (MoE) language models, released under the MIT license, deliver frontier-class performance across reasoning, coding, and agentic workloads while activating only a fraction of their total parameters per token, enabling high efficiency at scale. These models address different AI challenges with specialized capabilities: DeepSeek-V4-Pro is designed for complex reasoning and high-quality text generation at flagship scale. This 1.6T-parameter MoE model activates 49B parameters per token and supports a 1M-token context window, making it well-suited for long-document analysis, multi-step reasoning, and demanding generation tasks that require maximum model capacity. DeepSeek-V4-Flash delivers strong performance at significantly lower compute cost as an efficient 284B-parameter MoE model activating 13B parameters per token. It supports the same 1M-token context window as V4-Pro, making it ideal for latency-sensitive and cost-conscious production deployments across a wide range of language tasks. DeepSeek-V4.1-Flash provides next-generation agentic efficiency with native multimodal support. Built on a novel Causal Encoder-Decoder (CED) architecture with a 552B-parameter backbone, it activates just 8B parameters during prefill and 16B during decode, supporting a 1M-token context window and 384K max output tokens. Its Compressed Sparse Attention 2 (CSA2) and FP4 KV caching reduce the global KV cache footprint to roughly 1/4 of V4-Flash, delivering substantially improved cost efficiency for agentic workloads, coding, and vision tasks. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

DeepReinforce AI's Ornith-1.0-9B and Ornith-1.0-35B models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models are part of the Ornith 1.0 family of open-source, self-scaffolding agentic coding models released under the MIT license, achieving state-of-the-art performance among open-source models of comparable size on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo, and OpenClaw. These models address different agentic coding challenges with specialized capabilities: Ornith-1.0-9B is a lightweight 9B-parameter dense model post-trained on Qwen 3.5 9B, designed for efficient single-GPU deployment. With a 256K-token context window and 32 decoder layers, it is well-suited for individual developers and small teams running agentic coding workflows on a single GPU — including laptops with consumer-grade hardware. Its self-improving training framework uses reinforcement learning to jointly optimize scaffold generation and solution rollouts, enabling the model to discover better search trajectories and generate higher-quality code solutions. Ornith-1.0-35B is a 35B-parameter Mixture-of-Experts (MoE) model post-trained on Qwen 3.5, delivering stronger agentic coding performance while remaining deployable on a single 24GB consumer GPU. Like its smaller sibling, it employs reinforcement learning to jointly optimize the scaffolds that drive coding tasks and the resulting solution rollouts, discovering superior code trajectories. Its MoE architecture activates only a subset of parameters per token, providing a strong balance of quality and efficiency for production agentic coding workloads. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

Beam, the first open weight model offered by US-based Reflection, is now available in Amazon SageMaker JumpStart, the model hub in Amazon SageMaker AI. According to Beam, they operate at the frontier of open weight models in its class, with strong performance across reasoning, coding, and agentic benchmarks. With it, customers can now build advanced coding agents and automated workflows on a frontier open weight model while keeping control of their data and infrastructure and adapting the model to their business. Beam runs alongside other models that customers already use in SageMaker AI and scales to meet production demand with the same controls, audit trails, and monitoring built into every SageMaker AI deployment. Additionally, Beam was ported and optimized for AWS Trainium3, giving customers a high-performance, cost-efficient option for deploying and scaling the model from day one. “Reflection is delivering frontier AI that developers can actually control, not just rent through an API. Bringing Beam to SageMaker AI puts that capability in front of millions of developers on infrastructure they already trust, running in their own AWS account, on their own data, fully in their control," said Misha Laskin, co-founder and CEO, Reflection. Engineering teams can use Beam to resolve software issues, generate and review code, and automate routine operations work. Businesses can also use it for complex workflows that span many steps and tools, and for advanced math and science reasoning. Because Beam is released under the Apache 2.0 license, customers can inspect and customize the model, while also running it in their own AWS accounts, so data stays private and under their control, protected by the same AWS security and compliance capabilities they already rely on.  “Advanced AI models like Beam, now available on Amazon SageMaker, let us run open weight frontier AI entirely inside our own secure environments, with full visibility into the weights, the data, and the pipeline. In the age of multi-model AI, the pairing of openness and managed infrastructure is an essential addition to the AI ecosystem. The freedom to deploy and adapt advanced AI, tailored to mission problems with client-specific security and controls, gives more agencies greater flexibility to deploy advanced AI across their enterprise." said Garrett Berntsen, Chief AI Officer, Accenture Federal Services. Beam is initially available in Amazon SageMaker JumpStart in US East (N. Virginia) and US East (Ohio). To start, search for "Beam" in SageMaker JumpStart within Amazon SageMaker Studio, or deploy it directly on SageMaker AI.  To learn more, read the AWS/Beam launch blog, SageMaker AI documentation and SageMaker JumpStart documentation.

Beam, the first open weight model offered by US-based Reflection, is now available in Amazon SageMaker JumpStart, the model hub in Amazon SageMaker AI. According to Beam, they operate at the frontier of open weight models in its class, with strong performance across reasoning, coding, and agentic benchmarks. With it, customers can now build advanced coding agents and automated workflows on a frontier open weight model while keeping control of their data and infrastructure and adapting the model to their business. Beam runs alongside other models that customers already use in SageMaker AI and scales to meet production demand with the same controls, audit trails, and monitoring built into every SageMaker AI deployment. Additionally, Beam was ported and optimized for AWS Trainium3, giving customers a high-performance, cost-efficient option for deploying and scaling the model from day one. “Reflection is delivering frontier AI that developers can actually control, not just rent through an API. Bringing Beam to SageMaker AI puts that capability in front of millions of developers on infrastructure they already trust, running in their own AWS account, on their own data, fully in their control," said Misha Laskin, co-founder and CEO, Reflection. Engineering teams can use Beam to resolve software issues, generate and review code, and automate routine operations work. Businesses can also use it for complex workflows that span many steps and tools, and for advanced math and science reasoning. Because Beam is released under the Apache 2.0 license, customers can inspect and customize the model, while also running it in their own AWS accounts, so data stays private and under their control, protected by the same AWS security and compliance capabilities they already rely on.  “Advanced AI models like Beam, now available on Amazon SageMaker, let us run open weight frontier AI entirely inside our own secure environments, with full visibility into the weights, the data, and the pipeline. In the age of multi-model AI, the pairing of openness and managed infrastructure is an essential addition to the AI ecosystem. The freedom to deploy and adapt advanced AI, tailored to mission problems with client-specific security and controls, gives more agencies greater flexibility to deploy advanced AI across their enterprise." said Garrett Berntsen, Chief AI Officer, Accenture Federal Services. Beam is initially available in Amazon SageMaker JumpStart in US East (N. Virginia) and US East (Ohio). To start, search for "Beam" in SageMaker JumpStart within Amazon SageMaker Studio, or deploy it directly on http://aws.amazon.com/sagemaker/ai/.  To learn more, read the https://aws.amazon.com/blogs/machine-learning/introducing-reflection-beam-an-open-weight-model-for-agentic-coding-on-aws/, https://docs.aws.amazon.com/sagemaker/ and https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

NVIDIA's Nemotron 3.5 Super VL is now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. This vision-language model from the Nemotron 3.5 family accepts images, video, and text as input and generates text, enabling customers to deploy multimodal AI solutions on AWS infrastructure. Nemotron 3.5 Super VL is designed for document intelligence, GUI understanding, video question answering, and multimodal agentic applications. It pairs a 120B-parameter hybrid Mixture-of-Experts (MoE) language backbone with 12B active parameters and a vision encoder, supporting context lengths of up to 1 million tokens. The model includes a built-in Multi-Token Prediction (MTP) head for speculative decoding to accelerate text generation, with reasoning configurable per request. It supports ten languages including English, French, Spanish, German, Japanese, Hindi, Korean, and Chinese. Post-trained with multimodal supervised fine-tuning on image, video, document, chart, and GUI data followed by GRPO reinforcement learning, it delivers structured JSON output, tool calling, and configurable reasoning for enterprise agent workflows. SageMaker JumpStart offers two variants: NVFP4 for optimized inference on G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, and BF16 reference weights on ml.p5en.48xlarge for customization including SFT, RL, distillation, and domain adaptation. With SageMaker JumpStart, customers can deploy Nemotron 3.5 Super VL with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

NVIDIA's Nemotron 3.5 Super VL is now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. This vision-language model from the Nemotron 3.5 family accepts images, video, and text as input and generates text, enabling customers to deploy multimodal AI solutions on AWS infrastructure. Nemotron 3.5 Super VL is designed for document intelligence, GUI understanding, video question answering, and multimodal agentic applications. It pairs a 120B-parameter hybrid Mixture-of-Experts (MoE) language backbone with 12B active parameters and a vision encoder, supporting context lengths of up to 1 million tokens. The model includes a built-in Multi-Token Prediction (MTP) head for speculative decoding to accelerate text generation, with reasoning configurable per request. It supports ten languages including English, French, Spanish, German, Japanese, Hindi, Korean, and Chinese. Post-trained with multimodal supervised fine-tuning on image, video, document, chart, and GUI data followed by GRPO reinforcement learning, it delivers structured JSON output, tool calling, and configurable reasoning for enterprise agent workflows. SageMaker JumpStart offers two variants: NVFP4 for optimized inference on G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, and BF16 reference weights on ml.p5en.48xlarge for customization including SFT, RL, distillation, and domain adaptation. With SageMaker JumpStart, customers can deploy Nemotron 3.5 Super VL with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

Meta's Muse Spark Open is now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. This coding-optimized open-weight model from Meta is purpose-built for agentic coding workflows, bringing multi-step reasoning and tool calling to enterprise software engineering on AWS infrastructure. Muse Spark Open is designed for long-running coding agents and multimodal engineering workflows. Built on a Mixture-of-Experts (MoE) architecture with 816B total parameters and 42B active per token, it delivers the capacity of a very large network while keeping per-token compute closer to that of a much smaller dense model. The model holds a 1M-token context window, enabling it to ingest an entire project—including dependencies, legacy code, and thousands of files—without splitting tasks into smaller pieces. It accepts text, images, documents, audio, and audio-video as input and generates text output. Trained on real developer workflows such as multi-file refactors and long debugging sessions, it plans, acts, observes results, and self-corrects over long horizons rather than making single tool calls. SageMaker JumpStart offers two variants: BF16 (Muse-Spark-Open-816B-A42B) and NVFP4 quantized (Muse-Spark-Open-816B-A42B-NVFP4), with the NVFP4 variant requiring roughly a quarter of the memory for cost-efficient deployment. With SageMaker JumpStart, customers can deploy Muse Spark Open with just a few clicks, with the endpoint running in their own AWS account for full control over source code and prompts. To get started with this model, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the model to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

Meta's Muse Spark Open is now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. This coding-optimized open-weight model from Meta is purpose-built for agentic coding workflows, bringing multi-step reasoning and tool calling to enterprise software engineering on AWS infrastructure. Muse Spark Open is designed for long-running coding agents and multimodal engineering workflows. Built on a Mixture-of-Experts (MoE) architecture with 816B total parameters and 42B active per token, it delivers the capacity of a very large network while keeping per-token compute closer to that of a much smaller dense model. The model holds a 1M-token context window, enabling it to ingest an entire project—including dependencies, legacy code, and thousands of files—without splitting tasks into smaller pieces. It accepts text, images, documents, audio, and audio-video as input and generates text output. Trained on real developer workflows such as multi-file refactors and long debugging sessions, it plans, acts, observes results, and self-corrects over long horizons rather than making single tool calls. SageMaker JumpStart offers two variants: BF16 (Muse-Spark-Open-816B-A42B) and NVFP4 quantized (Muse-Spark-Open-816B-A42B-NVFP4), with the NVFP4 variant requiring roughly a quarter of the memory for cost-efficient deployment. With SageMaker JumpStart, customers can deploy Muse Spark Open with just a few clicks, with the endpoint running in their own AWS account for full control over source code and prompts. To get started with this model, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the model to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

Amazon SageMaker Studio now lets new customers sign up with their Google, GitHub, or Apple credentials and land inside a fully configured Studio environment - no traditional AWS account setup, no credit card, and no console navigation required. Previously, new customers who discovered a model on Hugging Face or an AWS product page had to complete a multi-step AWS account creation process, provide a credit card, navigate the AWS Console to find SageMaker AI, configure an environment, and set up IAM permissions before they could begin working. With earlier launches, we reduced environment setup to twenty seconds, pre-configured model customization permissions, and introduced deep links from Hugging Face and AWS product pages so customers could land directly inside Studio with a model pre-loaded. Now, with social sign-in, new customers complete a simplified sign-up using an existing identity, receive free credits as part of the AWS Free Tier with no credit card required, and arrive inside a ready-to-use SageMaker Studio environment where they can immediately fine-tune, evaluate, and deploy foundation models. Customers can invite team members by email, and collaborators sign in with their own social credentials to access the shared Studio environment, including shared models, notebooks, and endpoints without configuring IAM users or permissions. When a project grows beyond free credits, customers can upgrade to a paid plan and set a spend limit to stay within budget. Verified customers continue to receive default GPU access to G5, G6, G6e, and G4dn instances across endpoint deployments, training jobs, and notebooks without requesting quota increases. This feature is available in all AWS Commercial Regions where Amazon SageMaker Studio is supported. To get started, visit any supported model on https://huggingface.co/models and select "Customize and Deploy on SageMaker AI" or "Deploy on SageMaker AI," or click Get Started from the https://aws.amazon.com/sagemaker/ai/studio/ page. To learn more, see the https://docs.aws.amazon.com/accounts/latest/reference/sign-in-new.html and https://docs.aws.amazon.com/sagemaker/latest/dg/onboard-quick-start.html.

NVIDIA's Qwen3.6-35B-A3B-NVFP4 and Alibaba's Wan2.1-T2V-1.3B-Diffusers models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models bring specialized capabilities spanning agentic coding with long-context reasoning and lightweight text-to-video generation, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: Qwen3.6-35B-A3B-NVFP4 is optimized for agentic coding, multimodal reasoning, and long-context understanding as the NVIDIA-quantized variant of Alibaba's Qwen3.6-35B-A3B. This Mixture-of-Experts model contains 35B total parameters with only 3B activated per token (8 of 256 experts), supporting a 262K-token context window extendable to ~1M via YaRN scaling. Quantized to NVFP4 using NVIDIA's ModelOpt framework, it preserves thinking across conversation turns, multi-token prediction, and tool calling for multi-step agent pipelines—all at a significantly reduced memory footprint. Wan2.1-T2V-1.3B-Diffusers excels in text-to-video generation on consumer-grade hardware. Built on the diffusion transformer paradigm with a novel Video Variational Autoencoder (VAE), this 1.3B-parameter model generates high-quality, physics-consistent video clips from text prompts while requiring only 8.19 GB of VRAM. It can produce a 5-second 480p video on an RTX 4090 in approximately 4 minutes, making it one of the most accessible open-source video generation models available. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

BM's granite-speech-4.1-2b, Kakao's kanana-2-30b-a3b-instruct, and the OpenFold Consortium's OpenFold3 models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These three models bring specialized capabilities spanning multilingual speech recognition, bilingual agentic AI, and biomolecular structure prediction, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: granite-speech-4.1-2b is purpose-built for multilingual automatic speech recognition (ASR) and bidirectional speech translation (AST) across English, French, German, Spanish, Portuguese, and Japanese. This compact 2B-parameter speech-language model delivers a word error rate of 5.33% with a real-time factor of ~231, making it one of the most efficient ASR models in its class. Released under Apache 2.0, it integrates seamlessly into enterprise voice workflows for transcription, translation, and audio processing at scale. kanana-2-30b-a3b-instruct excels in bilingual Korean-English instruction following and agentic AI workflows. Developed by Kakao, it adopts a cutting-edge architecture featuring Multi-head Latent Attention (MLA) and Mixture-of-Experts (MoE), activating only 3B of its 30B total parameters per forward pass for superior throughput. Post-trained with supervised fine-tuning and reinforcement learning, it supports up to 128K tokens via YaRN scaling and is designed to function as an AI collaborator that understands context and acts proactively. OpenFold3 provides all-atom biomolecular complex structure prediction for proteins, DNA, RNA, and small-molecule ligands. Developed by the OpenFold Consortium and the AlQuraishi Lab at Columbia University, this diffusion-based model extends structure prediction beyond single proteins to model multi-chain complexes and heterogeneous biomolecular interactions. It supports computer-aided drug design and is applicable across academic and pharmaceutical research labs. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

Mistral AI's Ministral-3-3B-Instruct-2512 and Ministral-3-8B-Instruct-2512 models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models from the Ministral 3 family bring compact, vision-capable language models purpose-built for edge deployment and resource-constrained environments, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: Ministral-3-3B-Instruct-2512 is engineered for ultra-lightweight edge deployment with multimodal understanding. Comprising a 3.4B language model and a 0.4B vision encoder, it fits in just 8GB of VRAM in FP8 while supporting a 256K-token context window. It offers vision analysis, multilingual instruction following across dozens of languages (including English, French, Spanish, German, Chinese, Japanese, Korean, and Arabic), strong system-prompt adherence, and native function calling with structured JSON output—all under the Apache 2.0 license. Ministral-3-8B-Instruct-2512 delivers frontier-class capabilities comparable to its larger Mistral Small 3.2 24B counterpart in a compact 8B form factor. Built with an 8.4B language model and a 0.4B vision encoder, it fits in 12GB of VRAM in FP8 and features an interleaved sliding-window attention pattern for faster, memory-efficient inference. It shares the same vision, multilingual, agentic, and function-calling capabilities as its 3B sibling while offering stronger reasoning and generation performance. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

Google DeepMind's Gemma-4-31B-it-assistant and NVIDIA's Gemma-4-31B-IT-NVFP4 models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models bring the flagship Gemma 4 31B dense architecture to enterprise workloads in both full-precision and optimized quantized variants, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: Gemma-4-31B-it-assistant is built for multimodal reasoning, coding, and agentic workflows as the assistant-tuned variant of Google's flagship 31B dense model. It handles text and image inputs (including video as frame sequences) and generates text output, with a 256K-token context window and support for over 140 languages. Ranked #3 among open models on the Arena AI text leaderboard—outcompeting models 20x its size—it features a hybrid attention mechanism interleaving local sliding-window and full global attention with native function calling for building autonomous agents. Gemma-4-31B-IT-NVFP4 delivers the same Gemma 4 31B capabilities at a fraction of the memory footprint. Quantized with NVIDIA's ModelOpt framework to 4-bit FP4 precision, it reduces memory usage to ~18.5 GB (68% smaller than the base model) and achieves approximately 2.5x faster inference while retaining 97–99% of the original model's quality. Ideal for cost-efficient, high-throughput production deployments on NVIDIA RTX, DGX Spark, and data center GPUs. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

Mistral AI's Ministral-3-3B-Instruct-2512 and Ministral-3-8B-Instruct-2512 models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models from the Ministral 3 family bring compact, vision-capable language models purpose-built for edge deployment and resource-constrained environments, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: Ministral-3-3B-Instruct-2512 is engineered for ultra-lightweight edge deployment with multimodal understanding. Comprising a 3.4B language model and a 0.4B vision encoder, it fits in just 8GB of VRAM in FP8 while supporting a 256K-token context window. It offers vision analysis, multilingual instruction following across dozens of languages (including English, French, Spanish, German, Chinese, Japanese, Korean, and Arabic), strong system-prompt adherence, and native function calling with structured JSON output—all under the Apache 2.0 license. Ministral-3-8B-Instruct-2512 delivers frontier-class capabilities comparable to its larger Mistral Small 3.2 24B counterpart in a compact 8B form factor. Built with an 8.4B language model and a 0.4B vision encoder, it fits in 12GB of VRAM in FP8 and features an interleaved sliding-window attention pattern for faster, memory-efficient inference. It shares the same vision, multilingual, agentic, and function-calling capabilities as its 3B sibling while offering stronger reasoning and generation performance. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

BM's granite-speech-4.1-2b, Kakao's kanana-2-30b-a3b-instruct, and the OpenFold Consortium's OpenFold3 models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These three models bring specialized capabilities spanning multilingual speech recognition, bilingual agentic AI, and biomolecular structure prediction, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: granite-speech-4.1-2b is purpose-built for multilingual automatic speech recognition (ASR) and bidirectional speech translation (AST) across English, French, German, Spanish, Portuguese, and Japanese. This compact 2B-parameter speech-language model delivers a word error rate of 5.33% with a real-time factor of ~231, making it one of the most efficient ASR models in its class. Released under Apache 2.0, it integrates seamlessly into enterprise voice workflows for transcription, translation, and audio processing at scale. kanana-2-30b-a3b-instruct excels in bilingual Korean-English instruction following and agentic AI workflows. Developed by Kakao, it adopts a cutting-edge architecture featuring Multi-head Latent Attention (MLA) and Mixture-of-Experts (MoE), activating only 3B of its 30B total parameters per forward pass for superior throughput. Post-trained with supervised fine-tuning and reinforcement learning, it supports up to 128K tokens via YaRN scaling and is designed to function as an AI collaborator that understands context and acts proactively. OpenFold3 provides all-atom biomolecular complex structure prediction for proteins, DNA, RNA, and small-molecule ligands. Developed by the OpenFold Consortium and the AlQuraishi Lab at Columbia University, this diffusion-based model extends structure prediction beyond single proteins to model multi-chain complexes and heterogeneous biomolecular interactions. It supports computer-aided drug design and is applicable across academic and pharmaceutical research labs. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

NVIDIA's Qwen3.6-35B-A3B-NVFP4 and Alibaba's Wan2.1-T2V-1.3B-Diffusers models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models bring specialized capabilities spanning agentic coding with long-context reasoning and lightweight text-to-video generation, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: Qwen3.6-35B-A3B-NVFP4 is optimized for agentic coding, multimodal reasoning, and long-context understanding as the NVIDIA-quantized variant of Alibaba's Qwen3.6-35B-A3B. This Mixture-of-Experts model contains 35B total parameters with only 3B activated per token (8 of 256 experts), supporting a 262K-token context window extendable to ~1M via YaRN scaling. Quantized to NVFP4 using NVIDIA's ModelOpt framework, it preserves thinking across conversation turns, multi-token prediction, and tool calling for multi-step agent pipelines—all at a significantly reduced memory footprint. Wan2.1-T2V-1.3B-Diffusers excels in text-to-video generation on consumer-grade hardware. Built on the diffusion transformer paradigm with a novel Video Variational Autoencoder (VAE), this 1.3B-parameter model generates high-quality, physics-consistent video clips from text prompts while requiring only 8.19 GB of VRAM. It can produce a 5-second 480p video on an RTX 4090 in approximately 4 minutes, making it one of the most accessible open-source video generation models available. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

Google DeepMind's Gemma-4-31B-it-assistant and NVIDIA's Gemma-4-31B-IT-NVFP4 models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models bring the flagship Gemma 4 31B dense architecture to enterprise workloads in both full-precision and optimized quantized variants, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: Gemma-4-31B-it-assistant is built for multimodal reasoning, coding, and agentic workflows as the assistant-tuned variant of Google's flagship 31B dense model. It handles text and image inputs (including video as frame sequences) and generates text output, with a 256K-token context window and support for over 140 languages. Ranked #3 among open models on the Arena AI text leaderboard—outcompeting models 20x its size—it features a hybrid attention mechanism interleaving local sliding-window and full global attention with native function calling for building autonomous agents. Gemma-4-31B-IT-NVFP4 delivers the same Gemma 4 31B capabilities at a fraction of the memory footprint. Quantized with NVIDIA's ModelOpt framework to 4-bit FP4 precision, it reduces memory usage to ~18.5 GB (68% smaller than the base model) and achieves approximately 2.5x faster inference while retaining 97–99% of the original model's quality. Ideal for cost-efficient, high-throughput production deployments on NVIDIA RTX, DGX Spark, and data center GPUs. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

Meta's Muse-Glimmer-30B and Alibaba's Qwen 3.8-27B models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models bring specialized capabilities spanning autonomous local agentic workflows and multimodal long-horizon reasoning, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: Muse-Glimmer-30B is engineered for autonomous agentic tasks with multi-step reasoning, tool use, and failure recovery. This 30B-parameter dense model from Meta Superintelligence Lab combines a dedicated ~1.8B ViT-G/14 perception encoder with interleaved text and image inputs, a 131K+ context window, and selectable reasoning strength (low through extra-high). Released under Apache 2.0, it handles sequential tool calls, recovers from failures, and operates entirely without cloud infrastructure which is ideal for always-on enterprise agents. Qwen 3.8-27B excels in coding, multi-step agentic tasks, and multimodal understanding across text, images, and video. A dense 27B-parameter native vision-language model with a 262K context window (extendable to ~1M via YaRN scaling), it delivers substantial gains over its predecessor with adjustable reasoning effort levels. Scoring 61.7 on SWE-bench Pro and running at ~17GB quantized, it carries complex multi-step tasks through to completion with greater reliability. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

Meta's Muse-Glimmer-30B and Alibaba's Qwen 3.8-27B models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These two models bring specialized capabilities spanning autonomous local agentic workflows and multimodal long-horizon reasoning, enabling customers to deploy high-performance, scalable AI solutions on AWS infrastructure. These models address different enterprise AI challenges with specialized capabilities: Muse-Glimmer-30B is engineered for autonomous agentic tasks with multi-step reasoning, tool use, and failure recovery. This 30B-parameter dense model from Meta Superintelligence Lab combines a dedicated ~1.8B ViT-G/14 perception encoder with interleaved text and image inputs, a 131K+ context window, and selectable reasoning strength (low through extra-high). Released under Apache 2.0, it handles sequential tool calls, recovers from failures, and operates entirely without cloud infrastructure which is ideal for always-on enterprise agents. Qwen 3.8-27B excels in coding, multi-step agentic tasks, and multimodal understanding across text, images, and video. A dense 27B-parameter native vision-language model with a 262K context window (extendable to ~1M via YaRN scaling), it delivers substantial gains over its predecessor with adjustable reasoning effort levels. Scoring 61.7 on SWE-bench Pro and running at ~17GB quantized, it carries complex multi-step tasks through to completion with greater reliability. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

NVIDIA's Cosmos3-Edge, Cosmos3-Nano, and Cosmos3-Super models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These three models form the Cosmos 3 family of open, frontier omnimodal world models for physical AI, enabling customers to build robots, autonomous vehicles, and vision AI that perceive, reason, plan, and act in the physical world. These models address different physical AI challenges with specialized capabilities: Cosmos3-Edge is engineered for on-device robot control and real-time visual reasoning on edge hardware. This 4B-parameter omni-model (with a 2B Nemotron-based reasoner) operates at robot-control resolution (640×360), delivering real-time reasoning and generating 32 actions per inference at 15 Hz on NVIDIA Jetson Thor. It supports 256p and 480p video at 12–30 FPS, bringing frontier physical AI capabilities directly to embedded systems. Cosmos3-Nano excels in physics-aware world generation and physical reasoning as a compact 16B-parameter omnimodal model. It processes combinations of text, image, video, audio, and action trajectories to produce corresponding outputs, enabling robots and vision AI agents to reason using prior knowledge, physics understanding, and common sense. It supports chain-of-thought reasoning over text, images, and video with resolutions up to 720p. Cosmos3-Super provides the highest-fidelity world generation and simulation in the Cosmos 3 family at 64B parameters. It jointly processes and generates language, images, video, audio, and action sequences within a unified Mixture-of-Transformers architecture, supporting resolutions up to 720p across multiple aspect ratios. Ideal for large-scale simulation, synthetic data generation, and policy learning workflows. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the Amazon SageMaker JumpStart documentation.

NVIDIA's Cosmos3-Edge, Cosmos3-Nano, and Cosmos3-Super models are now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models available to AWS customers. These three models form the Cosmos 3 family of open, frontier omnimodal world models for physical AI, enabling customers to build robots, autonomous vehicles, and vision AI that perceive, reason, plan, and act in the physical world. These models address different physical AI challenges with specialized capabilities: Cosmos3-Edge is engineered for on-device robot control and real-time visual reasoning on edge hardware. This 4B-parameter omni-model (with a 2B Nemotron-based reasoner) operates at robot-control resolution (640×360), delivering real-time reasoning and generating 32 actions per inference at 15 Hz on NVIDIA Jetson Thor. It supports 256p and 480p video at 12–30 FPS, bringing frontier physical AI capabilities directly to embedded systems. Cosmos3-Nano excels in physics-aware world generation and physical reasoning as a compact 16B-parameter omnimodal model. It processes combinations of text, image, video, audio, and action trajectories to produce corresponding outputs, enabling robots and vision AI agents to reason using prior knowledge, physics understanding, and common sense. It supports chain-of-thought reasoning over text, images, and video with resolutions up to 720p. Cosmos3-Super provides the highest-fidelity world generation and simulation in the Cosmos 3 family at 64B parameters. It jointly processes and generates language, images, video, audio, and action sequences within a unified Mixture-of-Transformers architecture, supporting resolutions up to 720p across multiple aspect ratios. Ideal for large-scale simulation, synthetic data generation, and policy learning workflows. With SageMaker JumpStart, customers can deploy any of these models with just a few clicks to address their specific AI use cases. To get started with these models, navigate to the SageMaker JumpStart model catalog in the SageMaker console or use the SageMaker Python SDK to deploy the models to your AWS account. For more information about deploying and using foundation models in SageMaker JumpStart, see the https://docs.aws.amazon.com/sagemaker/latest/dg/studio-jumpstart.html.

最近次とOSSOpenSourceSoftwareの日本語LLMが発表されていますAma

本講演では近年のAI技術の進化とともに重要性を増している生成系AIモデルのファインチューニングに焦点を当てて解説します初めに生成系AIのファインチューニングの意義や利点を明らかにします次にAmazonSageMakerJumpStartを活用した簡便なファインチューニング手法について深堀りしますSageMakerJumpStartはAIMLのプロジェクトを迅速に開始するためのソリューションを提供し実際のビジネスシーンでの導入をスムーズにするためのツールやすぐ試せるモデルのカタログを備えています本セッションを通じて参加者はファインチューニングの実際の手法やAWSのサービスを活用した効率的なAIモデル開発のノウハウを習得することが期待されます登壇者本橋和貴アマゾンウェブサービスジャパン合同会社機械学習パートナーソリューションアーキテクト