- Parameters
- 3B
- Context
- 128k
- AgentBench-2
- 48.1
Runs comfortably on a single 24 GB accelerator. Intended for evaluation work, on-device tool use, and as a distillation target.
Models
Four checkpoints, one training recipe. Released weights for the Mini and Base sizes; the Pro sizes are available to research partners under a capability agreement.
Runs comfortably on a single 24 GB accelerator. Intended for evaluation work, on-device tool use, and as a distillation target.
Our general-purpose open checkpoint. Strong tool-use calibration and the smallest model that reliably completes long multi-step tasks.
Frontier checkpoint, shared with research partners under a capability agreement. Requires a documented review of intended use.
Trained for context rollover: state is summarised and re-expanded instead of truncated. Best results on tasks exceeding 500k tokens.
Every release ships with a card covering training data, evaluation, known limitations and the outcome of the pre-release review. The summary below is the short form; the full card is included with the weights.
| Release date | 14 August 2026 |
|---|---|
| Training data | Public web text to March 2026, permissively licensed code, and 1.4M synthetic agent trajectories generated in simulation. |
| Evaluation | AgentBench-2 61.7 · Terminal-2 54.9 · Tessera-information 0.62 · long-context retrieval 91.4% at 200k tokens. |
| Known limitations | Degrades on tasks requiring more than roughly 40 sequential tool calls. Over-uses the browser tool when a direct API call is available. Weak on non-Latin script tool outputs. |
| Pre-release review | Passed. Two red-team findings were fixed before release; one — a prompt pattern that induced unsafe file deletion in sandboxed mode — is documented in section 6 of the full card. |
| Licence | Apache 2.0 for weights and code. Use is subject to the Acceptable Use Policy. |