Language modelPreview
Agnes 3.0 Flash Preview
Agnes AI
- Released
- -
- Data date
- October 7, 2026
- Input
- Text, images, and video
- Output
- Text
The open Agnes 3.0 Flash Preview checkpoint has 33 billion parameters and an exact 262,144-token context. It accepts text, images, and video, returns text, and uses Apache 2.0.
Its hybrid architecture splits 72 decoder layers into 54 recurrent delta-rule layers and 18 global-attention layers. The model card explicitly separates this checkpoint from production Agnes 3.0 Flash; its benchmark results do not transfer to the preview weights either.
Specifications and access
| Specification | Value and source |
|---|---|
| Provider model ID | Agnes-AI/Agnes-3.0-Flash (preview weight checkpoint)Source |
| Access | Open weights for local deploymentSource |
| Input modalities | Text, images, and videoSource |
| Output modalities | TextSource |
| License | Apache 2.0Source |
| Architecture | 72 decoder layers: 54 recurrent delta-rule layers and 18 global-attention layers.Source |
| Checkpoint distinction | Earlier open-weight preview checkpoint; weights and benchmarks are distinct from the newer API checkpoint.Source |
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
IFBench (Show measurement, test conditions, and source)
- Source value
- 74.2
- Score
- 74.2
- Metric
- reported score; unit not stated in card
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Agnes-3.0-Flash Preview model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Provider's table for the open-weight Agnes-3.0-Flash Preview checkpoint only. Its rows combine results from different sources, harnesses and model snapshots; the card states these are not controlled head-to-head results. The table gives no exact model sampling settings or sample counts.
- Context
- These scores belong only to the earlier open-weight Preview checkpoint. The provider explicitly distinguishes it from the newer production/API Agnes 3.0 Flash model; no API scores are transferred here.
SciCode (Show measurement, test conditions, and source)
- Source value
- 38.08
- Score
- 38.08
- Metric
- reported score; unit not stated in card
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Agnes-3.0-Flash Preview model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Provider's table for the open-weight Agnes-3.0-Flash Preview checkpoint only. Its rows combine results from different sources, harnesses and model snapshots; the card states these are not controlled head-to-head results. The table gives no exact model sampling settings or sample counts.
- Context
- These scores belong only to the earlier open-weight Preview checkpoint. The provider explicitly distinguishes it from the newer production/API Agnes 3.0 Flash model; no API scores are transferred here.
GPQA Diamond (Show measurement, test conditions, and source)
- Source value
- 85.05
- Score
- 85.05
- Metric
- reported score; unit not stated in card
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Agnes-3.0-Flash Preview model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Provider's table for the open-weight Agnes-3.0-Flash Preview checkpoint only. Its rows combine results from different sources, harnesses and model snapshots; the card states these are not controlled head-to-head results. The table gives no exact model sampling settings or sample counts.
- Context
- These scores belong only to the earlier open-weight Preview checkpoint. The provider explicitly distinguishes it from the newer production/API Agnes 3.0 Flash model; no API scores are transferred here.
AA-LCR (Show measurement, test conditions, and source)
- Source value
- 68.33
- Score
- 68.33
- Metric
- reported score; unit not stated in card
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Agnes-3.0-Flash Preview model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Provider's table for the open-weight Agnes-3.0-Flash Preview checkpoint only. Its rows combine results from different sources, harnesses and model snapshots; the card states these are not controlled head-to-head results. The table gives no exact model sampling settings or sample counts.
- Context
- These scores belong only to the earlier open-weight Preview checkpoint. The provider explicitly distinguishes it from the newer production/API Agnes 3.0 Flash model; no API scores are transferred here.
AA-Omniscience Accuracy (Show measurement, test conditions, and source)
- Source value
- 23
- Score
- 23
- Metric
- reported score; unit not stated in card
- Unit
- No unit provided
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Agnes-3.0-Flash Preview model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. Provider's table for the open-weight Agnes-3.0-Flash Preview checkpoint only. Its rows combine results from different sources, harnesses and model snapshots; the card states these are not controlled head-to-head results. The table gives no exact model sampling settings or sample counts.
- Context
- These scores belong only to the earlier open-weight Preview checkpoint. The provider explicitly distinguishes it from the newer production/API Agnes 3.0 Flash model; no API scores are transferred here.
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Unresolved | preview-evaluation-protocol Original wording The Preview card publishes scores but no exact benchmark versions, sampling settings, model snapshots or sample counts. It notes that the rows come from different sources and harnesses. |
| Additional source | Model metadata (retrieved October 7, 2026) |