Release Notes
The release notes timeline highlights new features, bug fixes, and improvements implemented in Hyperstack.
The self-hosted Hyperstack API MCP Server is no longer supported. Use the hosted API MCP Server instead. It offers the same tools, there is nothing to install, and it signs you in with your Hyperstack account instead of an API key. To connect, see the Connection Guide.
Introduced Dedicated Inference, which runs an open-weight model of your choosing on GPUs reserved for your organization, reachable at a private HTTPS endpoint that only its API key can call. It is the per-hour counterpart to the shared per-token service, for predictable latency, a private endpoint, or a model the shared catalog does not serve.
- Deploy AI wizard: Select a model from the catalog or add a repository from Hugging Face, and the wizard creates everything needed to serve it: the GPU machine, the vLLM runtime, the model weights, the endpoint URL, and the API key. Start from Deploy AI Model on the Dashboard. Learn more.
- Model-based GPU sizing: GPUs are sized from the model's weights plus the key-value cache required by the context length and concurrency you set, under a Cost, Fast, or Experiment profile. Only configurations in stock can be selected, and you can override the recommendation with any configuration that fits. Learn more.
- OpenAI-compatible endpoint: Each endpoint serves chat completions at its own hostname, protected by the API key issued with it. Existing OpenAI-compatible clients work by changing the base URL and key. Learn more.
- Endpoint management: The Inference page lists each endpoint with its model, GPUs, region, status, and running cost, and links to the virtual machine that serves it. Deleting either one releases both. Learn more.
The Hyperstack API MCP Server is now hosted for you, so there is nothing to install and no API key to configure. The first time a client connects, your browser opens the Hyperstack sign-in page, and every tool call from then on runs against your own account.
- Connect any MCP client: Claude Code connects natively over HTTP with a single command, with no proxy and no packages to install. Claude Desktop connects natively as a custom connector, and Open WebUI connects as a tool server. Learn more.
- The full tool surface: Every operation the MCP Server already supported is available on the hosted server, covering virtual machines, Kubernetes clusters, volumes, environments, and billing. Learn more.
- Self-hosting: Running your own instance with your own API key is still supported, with its own installation and configuration guides. Learn more.
Fine-tuning and Hyperstack-hosted base models have been removed from the console and the API. As a result, the following features are also no longer available: LoRA adapter training and import, model deployment, model aliases, model evaluations, training metrics, logs, datasets, and data synthesis. AI Studio now offers inference exclusively on third-party hosted models.
- Retired models:
openai/gpt-oss-120b,mistralai/Mistral-Small-24B-Instruct-2501,meta-llama/Llama-3.3-70B-Instruct, andmeta-llama/Llama-3.1-8B-Instruct. Select a replacement on the Model Catalog page and update the model name in your requests. Learn more. - Unaffected: Inference through the API, the Playground, Knowledge Bases, system prompts, and conversations all continue to work as before. Billing history for past usage remains available.
Build a Knowledge Base from your own text and Markdown files on the new Knowledge Bases page, then select it in the Text Playground so models answer from your content instead of their training data alone. Learn more.
- Manage your files: Add and remove files at any time, follow indexing progress as it runs, and cancel indexing you no longer need from the Knowledge Base detail page.
- Cited sources: Grounded replies carry numbered citation markers. Hover over one to see the source file, an excerpt of the passage, and how relevant it was scored.
- API support: Manage knowledge bases programmatically with the new Knowledge Bases APIs, and ground a message by passing a knowledge base to the Send message endpoint.
- UI / API toggle in compare mode: In Compare Side-by-Side, the toggle now appears once, on the left panel, and applies to both panels when you switch between UI and API.
- Clearer errors for tool and function definitions: The Create chat completion API now returns a descriptive error when a request carries tool or function definitions it cannot process, in place of the generic failures that varied by model.
Kimi K3 is now available as a third-party model in AI Studio. Select it in the Text Playground, or call it through the API as moonshotai/Kimi-K3. The model accepts both text and image input, returns text, and supports a 1 million token context window.
Images in the Text Playground: Image-to-text models are now supported. You can attach an image to your message and have the model describe it or answer questions about it. Learn more.
- New image-to-text models: A catalog of image-to-text models is now available in the Model Catalog - Base Models, where you can filter by the Image-to-Text modality.
- API support: The Create chat completion endpoint now accepts multimodal message content for images.
Conversation history: Text Playground sessions are now saved to your conversation history. Browse, rename, and delete past conversations from the History tab, with full message history including any attached images. Manage conversations programmatically with the Conversations APIs.
- Clearer error messaging: When a transient backend issue interrupts a request, you now see a clear retry message instead of a generic credit-related error.
Image Playground: A new dedicated Image Playground is now available alongside the existing Text Playground. Generate and edit images using a catalog of image generation models (vision models), with customizable parameters and a side-by-side compare feature for evaluating multiple models at once. Learn more.
- Text-to-image generation: Generate images from a text prompt using any supported image generation model (text-to-image vision models). Learn more.
- Image-to-image editing: Upload a reference image alongside a text prompt to generate a modified or transformed result using image generation models (image-to-image vision models). Learn more.
- API view with cURL support: The Image Playground includes a UI / API toggle. Switching to API view displays the corresponding cURL command for your current request, adapting dynamically for text-to-image (Generate image from text) and image-to-image (Generate image from image) workflows. Learn more.
Vision models: A new catalog of image generation models (vision models) is now available, powering the Image Playground. You can browse and filter by modality (text-to-text, text-to-image, and image-to-image) in the Model Catalog - Base Models. The same modality filter is also available via the API using the
modalitiesquery parameter on List base models.Vision models pricing: A new Vision Models Pricing page is now available, covering pricing for text-to-image and image-to-image models. Learn more.
Image generation billing: Image generation is billed per 16×16 patch (1 patch = 256 pixels, or a 16×16 pixel region of generated output). A dedicated Image Generation Usage Report is now available under Billing - Resource Activity. Billing history is also available via the API. Learn more.
Gemma-4 model inference: Fixed an issue that prevented the
pearl-ai/Gemma-4-31B-it-pearlthird-party hosted model from loading for inference.Model Catalog icons: Added icons for new third-party models so each model displays its correct icon in the Model Catalog and Playground model selector.
Model pricing updates: Improved the reliability of automated pricing updates for third-party models.
Kimi model stability: Fixed an issue that prevented Kimi models from running successfully.
Qwen model compatibility: Resolved an issue where Qwen models could return errors when using default inference settings.
Improved API key access for team members: Fixed an issue where newly invited organization members were unable to use API keys until logging in via the UI.
Platform enhancements: Improvements to usability and overall platform experience.
Model Catalog third-party inference link: Added a “Learn more” link to the Model Catalog - Base Models page explaining third-party inference usage.
General UI improvements: Various interface improvements across AI Studio, including clearer GPU naming, improved model catalog visuals, and updated form text for better clarity and consistency.
Introduced the Hyperstack MCP (Model Context Protocol) Server, enabling natural-language infrastructure management through MCP-compatible AI clients such as Claude Desktop and Open WebUI. The MCP Server translates plain English instructions into secure, authenticated Hyperstack API operations, allowing users to create, manage, and monitor infrastructure resources without writing manual API calls.
Get started with our Quickstart Guide
Explore all MCP Server documentation
Third-party hosted models: AI Studio now offers selected third-party models for inference-only, accessible directly from the Playground. Learn more.
Persistent Playground chat: The Playground now supports persistent chat sessions, retaining conversation history, selected model, prompts, and parameter settings across page refreshes and logins.