Skip to main content

​
13th October 2026

​
Features

  • Improved support for open source models.
  • More models are available to agent ensembles. Run hive list models to see them.

​
Fixes

  • shmsize is rejected up front on providers that do not support it, instead of failing later.

​
6th October 2026

​
Features

  • New open models are available to agent ensembles: deepseek.deepseek-v4-pro, deepseek.deepseek-v4.1-flash, moonshotai.kimi-k3, xiaomi.mimo-v2.6-pro, xiaomi.mimo-v2.6-flash, zai.glm-5.3, zai.glm-5.3-flash and nvidia.nemotron-3-ultra. Run hive list models to see them.
  • Java syntax highlighting in the dashboard code view.

​
Fixes

  • Improved token efficiency of Hive agents.

​
30th September 2026

​
Features

  • Choose the regions your sandboxes run in with provider.regions. Set it to all to use every region the platform offers, for better GPU availability. Leave it unset to use the us (United States) region.
  • We ask for optional feedback when you stop an experiment, so you can tell us why.
  • The experiment list and overview now show who created each experiment.

​
Fixes

  • Experiments are rejected up front if their sandbox image does not exist, instead of failing later.

​
Deprecations

  • Prompt evolution is deprecated and no longer runs.

​
22nd September 2026

​
Features

  • Give the agents PDFs to work from with prompt.attachments — local files, directories (searched recursively for PDFs), or PDF URLs:
    Each PDF may be up to 20 MiB. See PDF attachments for more details.

​
Deprecations

  • prompt.arxiv_ids is replaced by prompt.attachments.

​
15th September 2026

​
Features

  • Choose the agents that generate code for your experiment. Either name a preset ensemble:
    or spell out your own weighted mix of models:
    Leave agents unset to keep the Hive’s default ensemble. hive list models lists the models an ensemble can name, and hive get exp shows the ensemble an experiment ran with. See agents.
  • Export an evolution report as a patch series: the report tab now offers a download that gives you one commit per idea, in report order, ready to apply to your repository with git am.
  • hive download repo supports large repositories. The zip is assembled server-side while the CLI polls and shows progress, so downloads no longer fail on a size limit. hive shell --exp EXPERIMENT likewise waits for the experiment’s code snapshot to be ready instead of failing when it isn’t.
  • Better error handling in the CLI and dashboard UI.
  • Install the Hive setup skill for your coding agent with hive skills install — it teaches the agent to write and configure hive.yaml, including multi-evaluators and GPU provisioning. Claude Code, Codex, Gemini CLI and Antigravity are supported.
  • hive create exp prints the dashboard URL for the new experiment.
  • Modal is available as a compute provider in beta. Talk to us if you would like it enabled for your team.

​
Fixes

  • Line ranges in target_code are now properly enforced, so code outside the range you asked to evolve is left alone.
  • Various dashboard and backend improvements.

​
Deprecations

  • The effort-tier coordinator configs are replaced by agent presets: coordinator_config_name: low-effort-config (and the medium- and high- equivalents) is now agents.preset: low-effort. The old names still work — they are rewritten to the equivalent preset — but setting both coordinator_config_name and agents is an error.

​
9th September 2026

​
Features

  • New evolution reports are generated periodically during an experiment, producing a summarized digest of the ideas and corresponding code changes which contributed most to the improved fitness.
  • Download the experiment data With Source: Only downloads the code of the files that have been changed by the Hive at each point, and not the entirety of target_code as before. Only applies to newly created experiments.
  • hive shell --exp EXPERIMENT creates an interactive sandbox containing the experiment’s code snapshot, see https://docs.hiverge.ai/gettingstarted/cli/reference#hive-shell.
  • Multi-evaluations, i.e. split one large evaluation into several smaller sub-evaluations that can be combined into a final fitness values. If the sub-evaluations can run on a smaller instance, e.g. a100:1 instead of a100:8, this can improve provisioning substantially. See https://docs.hiverge.ai/gettingstarted/cli/configuration#multi-evaluator
  • We improved Hive’s internal optimization and model use, and added support for newer models.
  • Better pagination of experiments.
  • You can now send feedback to us from the dashboard header.

​
Deprecations

  • runtime.num_agents is now called runtime.num_sandboxes in the experiment config (aka hive.yaml). The previous name num_agents internally set the number of sandboxes, not the number of agents. This change aligns the semantics, i.e. “number of sandboxes” with the name of the field, i.e. num_sandboxes. The correct migration is to simply rename the field in your config file. The old name is still accepted.

​
2nd September 2026

  • Exclude files from being evolved: repo.target_code=["!file.py"] allows the Hive to evolve every file except for file.py.
  • Improvements to the Results page, now primarily diff-based (clicking a point shows the diff for that iteration). Read more here.
  • Bugfix: Fixed issue where experiments were sometimes showing the wrong status.

​
21st August 2026

  • Each experiment maintains an internal knowledge base of what it has learnt.
  • Support for GPT-5.6 models.
  • Bulk-download experiments, including the metadata and sources for all iterations in a single JSONL.
  • Plots (and other parts of the UI) refresh automatically.
  • Improved utilization of compute resources.
  • More resilient experiments, and security updates across the platform.
  • Set default code view to show the unified diff with respect to the initial code version.

​
29th July 2026

  • Improvements to the experiment dashboard, including iteration and experiment statuses.
  • Faster experiment startup: repositories now upload in parallel with initialization, so experiments get going sooner.
  • Evolve any file type: experiments can now evolve files with any extension, not just a fixed set.
  • More resilient experiments.

​
17th July 2026

  • Support new command hive shell — Interactive shell access into hive workloads

​
14th July 2026

  • Improved experiment dashboard.
  • Improved multi-objective search and a new sampling strategy.
  • Use arXiv papers to guide experiments.
  • Pull down repositories from the CLI with the new hive download repo command.

​
2nd July 2026

  • Bring your own container images: New hive push, hive list, and hive delete image commands let you upload and manage custom images for your experiments.
  • Cost limits and budget controls: Set spending caps on your experiments and manage them directly from the dashboard.

​
24th June 2026

  • Redesigned dashboard with an improved experiment results page, plus better search and filtering.
  • Stop, cancel, and delete experiments directly from the dashboard.
  • Chat with your results — a new assistant to explore experiment outcomes.
  • Download code versions as a zip.
  • “Better than seed” filter to show only improved results, and a comparison mode for contrasting runs.
  • Upload and patch experiments using local source files (via the CLI).
  • Improved log viewer interface.

​
16th June 2026

  • New hive logs command to stream coordinator and sandbox logs, with --source, --worker, and --no-follow options.
  • hive stop replaces hive delete for ending experiments.
  • hive create exp now uses -c/--config to point at the config file (-f is deprecated but still works).
  • Experiment source can now be a remote git URL (https://, ssh://, git@) or a local directory, and the new repo.files field controls which files are uploaded. Source can also be omitted entirely when the base image already contains the code.
  • Updated the evaluator output format: print {"status": "success", "result": {"fitness": ...}} on success or {"status": "failed", "error": ...} on failure to the final line of stdout.

​
29th May 2026

  • Improved agent support.
  • Checkpoint support to save and resume experiments.
  • Improved log inspection support.
  • Improved experiment management.

​
15th May 2026

  • Observability improvements for tracking the experiment process.
  • Bug fixes and stability improvements.

​
22nd April 2026

  • First full release of the Hive platform