Blog
News from Nenyax.
Releases, product changes and the occasional longer story about what we learned building them.
Changelog
What shipped, release by release.
Every notable change to the SDK. The contract itself is versioned separately.
0.1.0a2
Unreleased
- Comparisons
- Paired-revision replay for comparison jobs, pinned to exact revisions, with errored and missing cases never turned into fabricated scores.
- Isolation
- Comparison workers honor sandbox boundaries and don't inherit runner credentials; owned processes are cleaned up on cancel.
- Packages
- Downloads select an exact revision; archive checksums are validated separately, including same-content reverts.
0.1.0a1
First public alpha
- Environments
- One contract and one call across step, rollout and task modes; drivers for Gymnasium, reasoning-gym, OpenEnv, NeMo Gym, Verifiers and Harbor.
- Judging and learning
- Judges from exact match to rubric LLMs; in-context, GRPO-family, DPO and Tinker learners; a held-out gate that reverts harmful updates.
- Infrastructure
- Process and Docker sandboxes with warm pools, 18 model providers, 3 environment hubs, tracking callbacks and SFT or DPO exports.
- Known limitations
- Weight-updating learners are not yet benchmarked at scale; NeMo Gym agent servers are not driven yet.
Follow along
Where news lands first.
Release notes
Every SDK release is announced with what changed and what was verified.
The changelog
Every notable change, release by release, with what's verified and what isn't yet.
Discord
Questions, previews and the people building Nenyax.
Questions
Good to know.
As changes land. Every release is listed in the changelog with what changed and what is verified.
Follow along.
Releases are announced in the Discord, along with questions and previews.