Language modelOpen weights
K2 Horizon MoVA 36B-A4B
Institute of Foundation Models
- Released
- September 3, 2026
- Data date
- October 7, 2026
- Input
- Text
- Output
- Text
K2 Horizon MoVA 36B-A4B complements the dense 32B model with a sparse architecture. Mixture-of-Value Attention routes attention through experts, while roughly 4 billion of its 36 billion parameters are active per token.
Specifications and access
| Specification | Value and source |
|---|---|
| Parameters | 36B total parametersSource |
| Active parameters | 4B parameters per tokenSource |
| Context window | 524,288 tokensSource |
| Input modalities | TextSource |
| Output modalities | TextSource |
| Documented training stages | Pretraining, four midtraining stages, and two SFT phases.Source |
| License | Apache 2.0Source |
Page 1 of 2
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.
τ³-bench Banking (Show measurement, test conditions, and source)
- Source value
- 26.8
- Score
- 26.8
- Metric
- success rate
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
Terminal-Bench 2.1 (Show measurement, test conditions, and source)
- Source value
- 58.6
- Score
- 58.6
- Metric
- pass rate
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
SciCode (Show measurement, test conditions, and source)
- Source value
- 38.9
- Score
- 38.9
- Metric
- score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
Humanity's Last Exam, without tools (Show measurement, test conditions, and source)
- Source value
- 25.2
- Score
- 25.2
- Metric
- accuracy
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
GPQA Diamond (Show measurement, test conditions, and source)
- Source value
- 80.8
- Score
- 80.8
- Metric
- accuracy
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
CritPt (Show measurement, test conditions, and source)
- Source value
- 2.1
- Score
- 2.1
- Metric
- score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
AA-LCR (Show measurement, test conditions, and source)
- Source value
- 66.3
- Score
- 66.3
- Metric
- long-context reasoning score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
AA-Omniscience Accuracy (Show measurement, test conditions, and source)
- Source value
- 18.8
- Score
- 18.8
- Metric
- accuracy
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
AA-Omniscience Non-Hallucination (Show measurement, test conditions, and source)
- Source value
- 69.2
- Score
- 69.2
- Metric
- non-hallucination rate
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- IFM model card
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. IFM-reported result for the exact MoVA 36B-A4B checkpoint. The card gives the task and score; it does not disclose a per-row sample count or full harness details.
AutomationBench-AA (Show measurement, test conditions, and source)
- Source value
- 26.10203199558987
- Score
- 26.10203199558987
- Metric
- success share
- Unit
- %
- Category
- independent-evaluator
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis model profile
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Artificial Analysis's independent measurement of the variant explicitly labelled the reasoning version on its profile. The component belongs to Intelligence Index v4.3.2; the embedded model object provides the exact raw value.
- Context
- The public comparison table rounds normalized shares for display; the stored score preserves the raw model-object value multiplied by 100. This observation belongs to the exact Artificial Analysis reasoning-version profile and is not a cross-model cohort reconstructed here.
Terminal-Bench 4.0 (Show measurement, test conditions, and source)
- Source value
- 0
- Score
- 0
- Metric
- success share
- Unit
- %
- Benchmark version
- 4.0
- Category
- independent-evaluator
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis model profile
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Artificial Analysis's independent measurement of the variant explicitly labelled the reasoning version on its profile. The component belongs to Intelligence Index v4.3.2; the embedded model object provides the exact raw value.
- Context
- The public comparison table rounds normalized shares for display; the stored score preserves the raw model-object value multiplied by 100. This observation belongs to the exact Artificial Analysis reasoning-version profile and is not a cross-model cohort reconstructed here.
SciCode (Show measurement, test conditions, and source)
- Source value
- 40.0462962962963
- Score
- 40.0462962962963
- Metric
- success share
- Unit
- %
- Category
- independent-evaluator
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis model profile
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. Artificial Analysis's independent measurement of the variant explicitly labelled the reasoning version on its profile. The component belongs to Intelligence Index v4.3.2; the embedded model object provides the exact raw value.
- Context
- The public comparison table rounds normalized shares for display; the stored score preserves the raw model-object value multiplied by 100. This observation belongs to the exact Artificial Analysis reasoning-version profile and is not a cross-model cohort reconstructed here.
Page 1 of 2
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. 10 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | per-run-protocol-and-sample-counts Original wording The model card does not disclose repeat counts or full harness details for most individual values. |
| Unresolved | excluded-non-target-checkpoints Original wording The pinned IFM family table contains values for intermediate checkpoint mid_4_final. Those weights are different from the final MoVA 36B-A4B target profile. The midtraining values were therefore excluded; final model-card results remain. |
| Additional source | IFM K2 Horizon release announcement (retrieved October 7, 2026) · IFM K2 Horizon announcement · Editorial description reviewed October 7, 2026 |
| Additional source | IFM model card (retrieved October 7, 2026) |
| Additional source | Artificial Analysis model profile (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |