v1.35
More predictable billing, more resilient Endpoints deployments, and clearer instance recovery.
Billing and credits
- Instance billing now stops as soon as an instance leaves the Running state. Time spent stopping or terminating is no longer charged.
- Organizations on contracts with a 100% discount now show an effective hourly price of €0.
- Endpoints deployments stop automatically when an organization’s credit balance reaches zero.
Endpoints
- Replicas stuck in Starting can now recover automatically, with a manual retry available when needed.
- Scaling down pauses surplus replicas instead of deleting them, making it faster to scale back up.
- Dashboard metrics now report 7-day request counts and input/output token counts more accurately. p95 latency and error rate are measured over a 24-hour window.
- GPT-OSS 20B now supports OpenAI-compatible tool calls and reasoning output.
- Model variants are now served under their own names, avoiding collisions between variants.
Instance reliability
- Instances that fail to start retry a limited number of times before moving to a clear error state that you can retry manually.
- Instances stuck while stopping or terminating are cleaned up automatically.
- Reboots now have their own recovery window and no longer fall back to a stop.
- Invalid instance actions now return clearer errors instead of generic server errors.
SSH connectivity
Every instance now exposes a stable SSH endpoint on port 22. The endpoint is reported in the instance’s events so it can be discovered consistently as the instance moves through its lifecycle.v1.34
Public API access, organizations and teams, the production launch of Hivenet Inference API, and major reliability improvements across Compute.
Public Compute API v1
The first public Compute API is now available.You can use eight documented endpoints to:- List and inspect instances
- Read recent instance logs
- Start, stop, and terminate instances
- List available regions
- Browse available hardware presets
Organizations and teams
Organizations are now available to all Compute users, making it possible to manage instances, teammates, and credits together.You can now:- Create an organization
- Invite teammates by email
- Preview, resend, and manage pending invitations
- Assign roles to organization members
- Use a shared organization credit balance
Hivenet Inference API
The Hivenet Inference API is now available in production, with OpenAI-compatible endpoints running on NVIDIA RTX 5090 hardware.Five models are available at launch:- Qwen 3.6 27B
- Qwen 3.6 35B
- Gemma 4 31B
- Gemma 4 26B
- GPT-OSS 20B
- Failed replicas restart faster.
- Model weights are loaded from a local cache at each location, reducing dependency on external downloads during startup.
- API keys associated with an organization are removed when that organization is deleted.
Shared dedicated servers
Private Compute locations can now be shared across multiple organizations.This allows several organizations, including separate teams belonging to the same customer, to use infrastructure reserved on the same dedicated server without requiring a separate location for each organization.Instance reliability
A broad set of improvements makes core instance management and networking more reliable, including fixes across instance lifecycle, recovery, and networking behavior.Faster support for instance issues
Hivenet Support can now perform additional recovery actions directly when troubleshooting an instance, including rebuilding an instance or changing its instance type.This reduces the need to escalate some recovery cases to the Compute engineering team and can help get affected instances working again sooner.Platform reliability
This release also includes infrastructure changes that improve monitoring, incident detection, deployment reliability, and how Compute manages regions as the platform expands.Public Compute API
Public Compute API documentationThe new Compute API documentation is now available for developers building scripts, tools, and integrations with Hivenet.The documentation covers:Manage existing Compute instances through the Hivenet API.
- Listing, inspecting, starting, stopping, and terminating existing instances
- Reading recent instance logs
- Listing supported region codes and available hardware presets
- Authentication, pagination, error handling, and rate limits
- Request and response examples for every documented endpoint
RTX 4090 fleet retirement
On August 8, 2026, Hivenet completed the retirement of its RTX 4090 Compute fleet as part of the transition to newer GPU hardware.RTX 4090 instances are no longer available for new workloads. Current GPU workloads should use the presets available in the Compute console, including NVIDIA RTX 5090 configurations.Existing RTX 4090 instances and their local data were removed when the retired nodes were decommissioned.Older tutorials, benchmark articles, and changelog entries may still mention RTX 4090 hardware when it is relevant to the configuration tested or available at the time. These references are historical and do not indicate current availability.See GPU types and sizes for current fleet guidance.The RTX 4090 fleet has been retired from Compute with Hivenet.
v1.32
More accurate instance availabilityInstance presets now reflect real-time capacity more accurately. Configurations that can’t currently be provisioned appear disabled instead of disappearing from the list.Sold-out and storage-constrained states are now more precise, reducing cases where an instance looked available but failed to launch.Inference improvementsFalcon-OCR is now available in the managed model catalog.Managed models now start faster and more reliably by loading weights from a local cache at launch instead of pulling them from an external endpoint.vLLM runtime behavior is also more stable, with pinned dependencies and cleaner shutdown handling.Storage capacity checks improvedCompute now checks available disk capacity more accurately before launch, including scheduling limits, disk reservations, and live replica state.Networking improvementsPer-instance network controls now apply to both VM and container instances, helping make network behavior more predictable and reducing noisy-neighbor variability.More accurate instance availability, Falcon-OCR inference, and more predictable networking.
v1.31
GPU VM reliability improvedMore reliable GPU VM restarts, cleaner billing, and easier instance access.
- Fixed several issues that could prevent GPU VMs from restarting, booting, or recovering cleanly, especially on Ubuntu GPU VMs and larger eight-GPU configurations.
- NVIDIA driver configuration now persists more reliably across package reinstalls and kernel upgrades, reducing the chance of GPU drivers breaking after a reboot.
- Stuck instances are now detected and recovered more reliably.
- Deleted instances now clean up related resources more consistently and are marked as terminated correctly.
- Deprecated Ubuntu and older image variants have been removed from the template catalog so unsupported builds can no longer be launched.
- Instance connection panels now show full access URLs for each open port, making it easier to open Jupyter and custom services.
0 in the Connectivity panel.Billing reliability improved- Fixed billing drift on stopped and errored instances.
- Compute now checks stopped instances regularly and closes any billing records that should no longer be active.
Additional infrastructure and monitoring improvements were added to support more reliable regional routing and faster issue detection.
v1.30.2
GPU VMsThis release improves GPU VM reliability, instance cleanup, SSH compatibility, and auto top-up stability.
- Fixed GPU driver issues that could prevent GPU VMs from starting correctly on supported operating systems.
- Prevented automatic OS updates from breaking GPU drivers after a reboot on GPU VM images.
- Fixed a boot issue that could affect 8-GPU VMs.
- Improved image tagging for more predictable deployments across environments.
- Fixed cleanup issues that could leave resources behind after an instance was terminated.
- Fixed an issue that could incorrectly show some instances as sold out in certain environments.
- Instances stuck in an error state are now automatically terminated and cleaned up.
- Fixed several edge cases affecting instance status updates and restart flows.
- Added ECDSA SSH key support.
- Enforced a minimum RSA key length of 2048 bits.
- Improved SSH connection performance for large file transfers.
- Improved the reliability of auto top-up credit handling for accounts with automatic billing enabled.
v1.30.1
InstancesImprovements to instance management, billing, and reliability, plus a few UI polish updates.
- Instances can now be terminated while starting.
- Stopped instances are now retained for 10 days instead of 5.
- Improved the instance experience across several flows.
- Fixed OS image name display.
- Added a transaction history view.
- Lowered the minimum custom top-up amount from €20 to €10.
- Sold-out states are now handled more gracefully.
- Improved SSH key validation.
- Updated UI text and labels across the product.
v1.25
Pricing at the time of this release:Lower GPU pricing, faster startup, and smoother custom networking.
- RTX 4090: 0.20 EUR/hour
- RTX 5090: 0.40 EUR/hour
- Custom TCP, UDP, and HTTPS port configuration for improved flexibility.
- Upgraded SSH documentation and in-app help link for easier setup.
- Ubuntu 24.04 LTS, CUDA 12.8, and pre-configured Hugging Face cache.
- Faster time-to-first-training and consistent environments across GPUs.
- Updated sign-up profiling questions for clarity.
- Last seen tracking for active users on Compute.
v1.24
New pre-packaged vLLM models added (total of 10 now available):Expanded vLLM model catalog, faster launch times, and improved stability.
- Meta Llama-3.1 8B Instruct
- Mistral Small-3.1 24B Instruct
- Llama 3.3 70B Instruct
- Mistral Small-24B Instruct
- Qwen-2.5 VL 32B Instruct
- GPT OSS 20B
v1.23
HTTPS services are now available.vLLM inference servers added.Improved instance flow for smoother setup and management.More connectivity options introduced.HTTPS support, new inference options, and smoother instance setup.