Roadmap
By November 14, 2026: agent-native research
Section titled “By November 14, 2026: agent-native research”- Arena v2 with full support for autonomous agent research: a research gym with MCP tools to submit, calibrate, train and score candidates on the dev split while validation and CORE stay sealed.
- Metered compute and research traces for every attempt, so research cost is disclosed automatically.
- Self-improving research loops: agents improve how they search, not only what they find.
- Leaderboards by researcher: which LLM and framework found what, at what research cost.
November 2026: scaling
Section titled “November 2026: scaling”- Extend the reference training recipe to 400B tokens and to larger models, with the same protocol discipline: pinned data, matched compute and a cost ledger.
Upcoming results
Section titled “Upcoming results”- Hybrids: DeltaNet, QSA and MoBA hybrids.
- Audits: independent run-directory audits for the looped and sparse runs; pilot and resume acceptance evidence for the published hybrids.
Their model cards are already in the observatory, marked “Results coming soon”. Each result is published through the same result pull request as any contribution.
Planned
Section titled “Planned”- Checkpoints and the tokenizer on the Hugging Face Hub.
- Port the Others implementation (MQA, GQA, MLA) into the current code.
Have an idea or want to help? Open a discussion.